“MACHINE LEARNING? YOU ARE GOING INTO A GREAT AREA.” – INTERVIEW WITH TRIPADVIOR’S JEFF PALMUCCI

http://dataconomy.com/machine-learning-you-are-going-into-a-great-area-interview-with-tripadviors-jeff-palmucci/

WHAT ADVICE WOULD YOU GIVE TO YOUNG PROFESSIONALS LOOKING INTO FINDING THEIR FEET IN DATA SCIENCE?

That is interesting. Because one of the biggest problems right now, I think in machine learning practices is finding good people. Hiring takes a significant part of my day and it is very hard to find good people especially people that know the techniques and understand them well enough to apply them. So if you want to get into this area, machine learning or data science, now is the great time to do it because the demand is high and the supply is low. And what you need to do is prepare yourself and educate yourself – do more than a couple competitions but take the courses and make sure that you work hard. Because one thing about interviewing in this field is that it is very quantitative. You can tell pretty much off the bat if somebody knows what they are talking about. My best advice to somebody going into this area is, first of all I’d say you are going into a great area, you are going to be in high demand but study hard, make sure you know the techniques.

Thank you for your time.

shhh top secret....don't read the following; its the roadmap.

1. 
   
2. New
   

   
   
   

   Always EXPECT multiples in this arena. They will generally, but not always use what they have on their benches. Some will have to drop out as they don't have the right bits and pieces to put together.
   
   So INVN will be a multiple. Expect it. SO what are they planning on?
   
   
   
   

   Data Scientist (Job ID: 12443)
   

   • Leverage machine learning techniques to discover data patterns and user behavior
   • Evaluate scenarios, and predict future outcomes through statistical data modeling, big data, and optimization tools and techniques like Bayesian modeling , MCMC based estimation methods, convolution-based methods, machine learning, random forests,decision trees, etc
   • Use a combination of R, Python, JavaScript, C, Java, , SQL, etc to solve problems and discover new solutions
   • Translate ideas and theory into commercial solutions, while taking ownership of the process.
   • Help shape our data infrastructure, reporting and analytics platforms

   Requirements:
   

   • 8+ years in the field of data science, data discovery and machine learning.
   • Strong math background
   • Familiarity with a variety of machine learning techniques and statistical methods
   • Proficiency with machine learning techniques
   • Strong DB skills are a must, should be very strong with SQL and NoSQL databases
   • Proficient in Hadoop, Map/Reduce, Pig, Scala, Hive
   • Experience with designing and building large scale data pipelines at large scale
   • Extensive Linux systems/shell programming experience
   • Expertise in distributed/scalable systems and algorithms with awareness of time and space complexity
   • Experience with handling and mining geospatial data
   • Masters degree in Computer Science or equivalent

   jfieb, 5 minutes agoEditDeleteReport

   #7Reply
3. 
   
4. 
   
5. 
   
6. 
   
7. 

   jfiebMember

   New
   

   
   
   

   Compare that one to this one....
   
   Posted On10/26/2015
   Principal, Algorithm Architect
   
   
   Employment Type :Full Time Regular
   
   Job ClassProfessional
   The principal sensor algorithm architect plays a key role indeveloping advanced sensor algorithms on QuickLogic’ssensor hub for smartphone and wearable. Strong algorithm development andsoftware implementation experience, with the ability to collaborate withmulti-disciplined engineering teams, is required for success in this position.
   
   Inthis role, he or she is responsible for developing, enhancing and maintaining advancedsensor algorithms including motion sensor calibration and fusion, contextawareness, gesture recognition and other sensor algorithms on QuickLogic’ssmart-sensor hub for smartphones and wearable. He or she is responsible for allphases of development including generating algorithm specifications based on keyperformance indicators and other market requirements,algorithmsimulation, real-time implementation using C and Assembly, performance tuningfor various use cases, optimization for memory footprint and computation load.
   
   He or she is also responsible for guiding thetesting engineer for generating the testing methods and apparatus for sensoralgorithms, sensor data collection, visualization and annotation for sensor algorithms.
   
   The engineer will support for key customer engagement toensure design wins and successful integration, qualification and deployment ofthe smart sensor products. He or she may also engage with our partners,customers, and other engineering teams for generating the technical requirementsand generating the system and algorithm specifications
   
   Skills & Knowledge Profile
   
   

   • PhDwith 7+ years in electrical engineering or computer science is preferred. MSwith 10+ years in electrical engineering or computer science is required.
   • Expertlevel in algorithm development using sensor fusion, machine learning, patternrecognition, simulation and mathematical analysis. Familiar with algorithmdevelopment tools such as Matlab and Simulink.
   • Experiencein efficient implementation of floating point and fixed point algorithms in firmware,software and hardware.
   • Demonstratedability to successfully complete complex R&D technology projects
   • Mustbe able to work in team environment

   
   The following experience is highly desired:
   
   

   • Androidsystem architecture experience

   
   
   
   What can be discerned...
   
   Both are racing down their visions of what they see. Both will add intelligence.
   
   The current difference?
   
   INVN is trying to make a jump. A big jump.
   A jump where?
   
   They are going to try to move from the device into the cloud...machine learning in the cloud.
   
   
   

8. 1. 
      
      QUIK on the other hand currently sees their roadmap, with the Sensory work fresh under their belts, intelligence on the device, leave the Hadoop sort of BIG/massive data mining to others.....PS this past yr I spent several days reading about Hadoop, partly to find a way to invest in it.  Its use ?  Ability to handle just massive data that is required for machine learning in the cloud.  Came away with nothing,

   Sensory several yrs ago decided to focus on audio ON the device and NOT just in the cloud.

   Local intelligence with audio as a mental model, will be what QUIK can differentiate with.

   
   

   Local intelligence ( like audio ), low power, will be their  card. It can be done again and again.

   And IF it becomes ubiquitous?  That is a future hardcode engine.

   
   
   

Dr Saxe is in Scottsdale now,

Designing Wearable Technology for Maximum Flexibility, Performance, and Power Efficiency

Moderated by Tim Saxe, Chief Technology Officer, QuickLogic

A key challenge for software and hardware designers of wearable electronics is to provide advanced processing capabilities in a small form factor while minimizing power consumption. Given that wearable technology is still an emerging market, it is also critical that developers maintain flexibility to meet changing market needs. For example, devices must be able to support advanced signal processing capabilities to reliably extract sensor data from low level signals in noisy environments. Devices must also be able to adapt to new types of peripherals such as motion sensors woven into clothing as well as implement increasingly sophisticated algorithms that can determine, for example, whether a person is running, bicycling, or cross-training. This session will discuss the use of hybrid device architectures and explore how developers can combine software programmability with application-specific accelerators and reconfigurable hardware to achieve the performance, power efficiency, and integrated functionality needed to enable next-generation devices ranging from low-end activity monitors to high-end sports watches.

If anyone in the forum is at this great confab, please check it out for us.

THanks in advance.

That Dr. Saxe, he in demand now.

It is a coffee house and lots of things get discussed with lots of noisy conversation.  QUIK has things to offer the sensory coffee house and will also listen all the time to tweak the roadmap.

They will find where advanced algos are needed.

Please skip if you only want street sort of material.

QUIK is headed into waaay cool rooms adjacent to the ones they have just entered.

1. 
   
2. New
   

   
   
   

   QUIK is headed into waaay cool rooms adjacent to the ones they have just entered.
   
   Background material
   
   
   PATTERN RECOGNITION & SENSORY MEMORYON is the ability to identify objec the environment--a necessary first step in all cognitive processses. Incoming sensations are combined/compared with patterns stored in LTM.
   

   • It is related to SENSORY MEMORY, where incoming stimuli are held for further processing following their initial detection.

   TEMPLATE-MATCHING MODELS
   

   • Template: exact internal representation of a pattern to be recognized.
   • Incoming patterns are compared "as is" to existing, stored patterns in LTM.
   • Example: a computer scanning the bottom numbers on checks looks for an exact match, or the use of UPC codes.
   • Template-matching theory as a theory of pattern recognition:
   • 1.) assumes that a retinal image is faithfully encoded in the brain.
   • 2.) assumes that an attempt is made to compare the retinal image directly to various stored patterns, or templates.
   • 3.) may be a mechanism for the sensory register--allows extra time to hold information for processing, but with a questionable application for real life--when do we have stimuli changing every 250 ms on a consistent basis?
   • Problems with template matching theories:
   • 1.) Requires an all or none judgment.
   • 2.) Gives no room for context.
   • 3.) Symbols must always be the same size, orientation position as the pattern to be identified.
   • 4.) Things can go wrong with the template--assumes a precise retinal image.
   • 5.) Does not show how two patterns may vary, e.g. P vs R, or E vs F
   • 6.) Does not allow for alternate descriptions of the pattern, e.g., sting ray versus sail--ambiguous stimuli.
   • 7.) That a precise, standardized system is needed for template matching to work reduces the credibility of this process as a model for human pattern recognition, which is highly complex and would require an almost infinite store of patterns.
   • 8.) Very few researchers today view template matching as an adequate model.

   FEATURE ANALYSIS MODELS
   

   • Feature analysis as a theory of pattern recognition:
   • 1.) Requires the recognition of critical features rather than reading an exact template.
   • Stimuli are thought of as combinations of elemental features. Can describe a pattern by listing its parts.
   • Like bottom-up processing in that there is a feature-by-feature comparison. Individual units are used to build a whole.
   • 2.) Key component: contrasts/ differentiates between patterns and looks for distinctive features, i.e., the lowest horizontal line in E, differentiate it from F.
   • Advantages over template matching:
   • 1.) Features are simpler units.
   • At the visual level there is strong evidence that the nervous system indeed extracts such features as lines (e.g., Hubel & Weisel, 1962).
   • 2.) "Critical" features, those relationships among features which are most critical to a pattern, can be identified: i.e., for the letter ‘A’ the critical point is that two approximately 45-degree lines intersect as near to the top as possible and the cross bar intersects both lines as nearly as possible, bisecting both lines.
   • 3.) Reduces the number of stored units: you do not need a template for each possible pattern but only for each feature. Since the same features tend to occur in many patterns this would mean a considerable savings in storage.
   • 4.) Sensitivity to context: small variations in relationships can be overcome by taking account of the context--top-down processing is taken into account-- expectations and internal schemata play a part in perception.
   • 5.) Problem: Fails to explain relationships between features.

   STRUCTURAL THEORIES
   

   • A.) Extend feature theories by specifying how features are related. This is a follow-up step to feature identification.
   • B.) Also called ‘analysis-by-synthesis’.
   • C.) Makes use of Gestalt principles--rules for initial form organization. The Gestalt principles seem to function so that one can quickly segment a stimulus into a set of objects and to organize these objects into larger configurations.
   • In some situations this segmentation seems quite successful: i.e., in a written page the principle of proximity serves to identify the letters and words, the principles of proximity and good continuation identify the line of text.
   • Theoretically more efficient than template models: They allow a single set of rules for determining relations among features to apply to a wide range of objects and scenes.

   IMPORTANCE OF CONTEXT OR TOP-DOWN PROCESSING
   Context:
   
   

   • 1.) The use of existing knowledge to guide processing.
   • 2.) At the earliest, top down processing can take place at the level of visual or auditory short term memory.
   • 3.) Top-down processing becomes more important as you degrade bottom-up. We compare what we see with past experiences and use this to interpret what is being seen.

   Auditory example of verbal context
   effect on pattern recognition
   

   • Have you seen the new display?
   • Have you seen the nudist play?

   SENSORY REGISTERS; VISUAL & ACOUSTIC STORES: OVERVIEW
   

   • There appear to be sensory registers, or stores, which can hold incoming stimulus information for very short periods of time. These sensory registers are a first step in establishing a more permanent memory.
   • Each sensory system has its own, apparently independent store:
   • Iconic (vision) Auditory (hearing)
   • Haptic (touch) Gustatory (taste)
   • Olfactory (smell)

   THE INFORMATION AVAILABLE IN BRIEF VISUAL PRESENTATIONS
   

   • Brief visual information is gathered by use of a tachistoscope--stimuli can be viewed for a precise duration and a brightness.
   • A.) The Span of Apprehension
   • the amount of information we can attend to at any one time. Defined as 50% accuracy in detecting the information available in a visual display. This amount is limited, and the limit can be assessed by using a t-scope.
   • B.) The Partial Report Procedure
   • Sperling (1960) devised a procedure in which letters and numbers appeared in a 4 x 3 matrix. In a control, whole report condition, stimuli were exposed and subjects reported as many letters and numbers as they could--average was between 4 & 5.
   • This suggests there is a limited ability which is linked to a perceptual limit.
   • In the experimental, or partial report condition a tone of high, medium or low pitch was sounded, cueing the subject to report the stimuli from either the top, middle or bottom row.
   • .......................................

   ECHOIC MEMORY
   

   • A.) Dichotic Listening
   • The auditory sensory store has been inferred by the technique of dichotic listening.
   • Subjects wear stereo headphones through which different information is presented to each ear, simultaneously. If you present three digits to the left ear and three digits to the right ear simultaneously, then ask for recall, subjects typically recall first all of the digits from one ear and then all of the digits heard in the other.
   • This implies that the second set must be held while the first is being recalled.
   • Another dichotic technique involves shadowing, where a subject is presented two different stories, one to each ear, and the subject is asked to repeat aloud the story in one ear.
   • When the two messages are identical, but offset, subjects become aware of that fact when they shadow the leading message, with the nonshadowed message offset by up to 13 words. When the shadowed message is the lagging one, then subjects only become aware of them being identical when they are offset by up to 6 words. This is evidence that there must be some brief storage for the nonshadowed message as well.
   • B.) Partial Report
   • Darwin, Turvey & Crowder (972) applied Sperling's partial report technique to the auditory modality.
   • They found that by cuing recall with a visual marker the partial report was superior to the whole report.
   • For whole report the span of auditory apprehension is between 4 and 5 items, just as for the visual store. However, by partial report auditory apprehension increased, even for delays up to 4 secs, although the increase was not as large as for the visual technique.
   • The findings suggest there is an auditory store in which echoic memory persists longer than iconic memory, but which is more limited in its capacity.
   • ..........

   
   Many of these sort of items are not in the places I usually find tidbits; they are at Universities, and neuro departments. QUIK will move into the above in a gradual fashion, but there really is NOT end for a roadmap and it goes pretty far into the next generation of computing
   if we are lucky.
   
   Just where can you invest in such technology? Try to find one that is traded, not venture, that has a real basis( not vapor ).
   I keep looking

tiers of sensor intelligence

Discussion in 'Main Forum' started by jfieb, Oct 29, 2015.

Thread Tools

Unwatch Thread

1. 
   

   A long term project and can be skipped as a digression.
   
   
   Ricks Churhill quote.
   
   
   “Now this is not the end it is not even the beginning of the end but it is, perhaps the end of the beginning.” -Winston Churchill
   
   A great quote that works very well. Consider sensor fusion intelligence.
   
   
   I have used the rice terreaces as a mental model for several yrs now, but the same quote applies.
   Rick's nice statement in the software segment speaks of the margins of software IP-how good they are.
   
   We have an MCU now above the FFE. QUIK will use it to the MAX for intelligence.
   
   Consider that sensory is ahead of the pack by one major decision they made yrs ago.
   Keep audio local, deeply embedded on the device and then it becomes a UI even if you have no connection.
   That meant they focused on power well before most others. It is a great fit- QUIK and sensory.
   
   
   On math units... they are not equal, I won't go in to all the headache inducing details, but a good yrs of reading told me the following....
   
   Kalmans ( a kind of math ) filters are the basis of sensor fusion, to be low power you need a math unit and so most, but not all, sensor fusion hubs use the M4 for its floating point math unit......BUT Kalmans- if given an ideal- use FIXED point math unit. So even though no one told me, the technology told me that the FFE is a fixed point unit made for running Kalmans, BETTER than an M4 floating point.
   
   The EOS then has 2 - count em math units, one for basic fusion in the FFE and the FP unit in the MCU.
   A run of the mill MCU has the floating point only, not 2 math units as the Eos must have.
   
   
   To maintain margins QUIK will layer on intelligence just like the rice terraces. So this thread will be my efforts to explore what these layers of intelligence might look like.....
   
   this is just to plant the seeds that the roadmap will be very interesing as it is the beginning of intelligence above the Fusion- expect that QUIK is ON this already. A lot will be done in the cloud, but audience has shown us that there can be real benefit for intelligence that still resides ON the device.
   
   I will be looking for that- more intelligence that resides on the device, and am using the audio engine as a mental model.
   
   So if this is something thatsounds interesting, read along.
   This is how I have fun as a part owner of the QUIK biz.
   If you find it gives a furrowed brow, or causes a headache it is not mandatory reading
   & can be skipped, while waitng for what sounds like a very nice CES.

    

   Last edited: Saturday at 3:06 AM

   jfieb, Oct 29, 2015EditReport

   #1Reply

   John likes this.
2. 

   jfiebMember

   
   
   
   
   
   QUIK will add layers of intelligence on top of what we have now.
   Sensory shows us the benefit of keeping some intelligence local.

    

   jfieb, Saturday at 3:11 AMEditDeleteReport

   #2Reply
3. 

   jfiebMember

   Sensory shows us the benefit of keeping some intelligence local
   
   
   Local machine learning.....on sensor data!
   
   I was uncertain if we could keep it local
   Android Marshmallow adds 'high fidelity sensor support' flag for developers
   

   Google is empowering developers with new tools to better deal with wonky sensors, which have long been a sore spot of Android fragmentation.
   
   According to the Android 6.0 Compatibility Definition Document, devices whose sensors are accurate to within very strict tolerances can set a new flag: android.hardware.sensor.hifi_sensors. Devices whose accelerometer, gyroscope, compass, barometers, step detectors, etc. all deliver data with high accuracy and broad range must set this flag. This is a boon for developers, who can look for a single value and know it can rely on the sensor data being accurate (or at least, put up a warning message to users that their device might deliver a sub-par experience).
   
   Currently, developers can look at various flags to determine if a device as a particular sensor or not, but they have no way of knowing if it delivers precise, low-latency data.
   
   The Android compatibility document also lays out power requirements so hardware manufacturers can build in sensors that will work as Google intends. Devices don't have to meet the new requirements, they're entirely optional, but the existence of a standardized way to tell developers they can rely on sensor data being accurate and fast will be a big help in cutting down one of the prominent pain points in Android fragmentation.
   
   Why this matters: One of the reasons certain apps appear first, or only, on iOS is because iPhones deliver very consistent sensor data from one model to the next. Android phones, in being so diverse, often give wildly different results. What's more, developers can't rely on the results being delivered in a timely, low-latency fashion. Google is making it easier, if optional, for developers to know if a device provides high-quality, reliable, fast sensor data. In time, you may start seeing applications that are especially sensor-dependent (like running trackers or motion-sensitive games) throw up a warning on devices that don't set the android.hardware.sensor.hifi_

    

   jfieb, 20 minutes agoEditDeleteReport

   #3Reply
4. 

   jfiebMember

   Can be skipped, but I want a solid foundation.
   For what some ask?
   
   For EOS 2. The intelligence goes up
   
   
   STANFORD UNIVERSITY
   Machine Learning
   
   About this Course
   Machine learning is the science of getting computers to act without being explicitly programmed. In the past decade, machine learning has given us self-driving cars, practical speech recognition, effective web search, and a vastly improved understanding of the human genome. Machine learning is so pervasive today that you probably use it dozens of times a day without knowing it. Many researchers also think it is the best way to make progress towards human-level AI. In this class, you will learn about the most effective machine learning techniques, and gain practice implementing them and getting them to work for yourself. More importantly, you'll learn about not only the theoretical underpinnings of learning, but also gain the practical know-how needed to quickly and powerfully apply these techniques to new problems. Finally, you'll learn about some of Silicon Valley's best practices in innovation as it pertains to machine learning and AI. This course provides a broad introduction to machine learning, datamining, and statistical pattern recognition. Topics include: (i) Supervised learning (parametric/non-parametric algorithms, support vector machines, kernels, neural networks). (ii) Unsupervised learning (clustering, dimensionality reduction, recommender systems, deep learning). (iii) Best practices in machine learning (bias/variance theory; innovation process in machine learning and AI). The course will also draw from numerous case studies and applications, so that you'll also learn how to apply learning algorithms to building smart robots (perception, control), text understanding (web search, anti-spam), computer vision, medical informatics, audio, database mining, and other areas.
   
   In what form will local (/deeply embedded ) learning occur on a mobile device.
   
   Sensory shows us the benefit of keeping some intelligence local
   
   + Dr. Saxe snip of text.
   
   Whats good ?
   
   Machine learning will be a margin maintaining adjacent possibility. It a room that QUIK gets to explore starting now in quiet mode, with some things to read in the coming months.

    

   jfieb, 13 minutes agoEditDeleteReport

   #4Reply
5. 

   jfiebMember

   New

   meanwhile later this month
   
   “An Ounce of Preprocessing Is Worth a Pound of Computing”
   
   Wearable Sensors and Electronics 2015
   Annual Conference and Exhibition
   November 16 – 17, 2015
   Santa Clara, California
   
   Conference Topics
   Who Should Attend
   Conference Location
   Hotel Information
   Testimonials
   Contact Information
   Sponsorship Opportunities
   Exhibit Opportunities
   Registration
   Click to view participating companies
   
   
   This unique event has an exclusive focus on sensors and electronics for wearable applications. Following the PC and smartphone waves of development, wearables are the "next big thing". And what makes various wearable designs unique is the ability to pack advanced sensors and electronics into very small form factors. Sensors are truly the enabling technology for these applications -- wearables simply cannot exist without sensors.
   
   So what are the next-generation wearable sensing technologies? And what are the applications that are driving the need for new types of sensors? These are the key questions that will drive the discussion at this second annual conference, Wearable Sensors and Electronics 2015, which will feature talks from the leading wearable sensors and electronics technology experts. The future is bright for wearables -- attend this event to identify emerging technology and application trends, exchange ideas, form new companies, and network with your industry peers!
   
   Conference Topics
   
   

   • Worldwide wearable device trends: market drivers, demographic factors, emerging markets and applications, disruptive technologies, government policy effects.
   • Business aspects: competitive forces and dynamics, pricing trends, mergers and acquisitions, analyst forecasts and projections, manufacturing developments, technology transfer, regulatory compliance, ecosystems and hubs, company formation.
   • Technology trends and developments: wearable device architecture, sensor hubs, ultra-low power systems and components, energy harvesting, micro batteries and energy storage, supercapacitors, sensor fusion, software algorithms, context awareness, virtual sensors, connectivity with smartphones.

   No wonder QUIK is going to be there?
   

   • Emerging applications: digital health, body area networks, medical diagnostics and screening, genomics, safety and security, environmental, virtual reality, indoor navigation, quantified self, usage paid insurance.
   • Emerging manufacturing techniques and materials: flexible and printed electronics, smart glass, streamlined assembly techniques.
   • Emerging types of sensors: touch, pressure, thermal, radiation, humidity, chemical, high- performance image and IR, air and pollution, magnetic, water, radar, high performance inertial, high performance microphones, microphone arrays.
   • Emerging types of actuators: high performance micro speakers, optical zoom, micro shutters, energy harvesters.

   Who Should Attend
   

   • CEOs
   • CTOs
   • Entrepreneurs
   • VPs of engineering
   • VPs of marketing
   • VPs of business development
   • MEMS foundry managers
   • Product managers and engineers
   • Marketing, sales and business development managers
   • MEMS and sensors production managers and engineers
   • MEMS and sensors design managers and engineers
   • Business development professionals
   • Industry analysts
   • Investment bankers
   • MEMS and sensors consultants
   • Media representatives