Eran Briman
Transcription
Eran Briman
Smart and Connected: Can Machines Exceed Humans? ChipEx 2016, Tel Aviv May 9, 2016 Quiz: What’s Common for: CEVA Proprietary Information 2 Connected Machines: Is it That Simple? Multiple and constantly evolving communication standards The IoT is composed of an almost endless list of comms standards 80 Others 70 802.15.4 60 GPS/GNSS Cellular (Incl. M2M) RFID (Active and Passive) Wi-Fi Bluetooth Units, Billions NFC 50 40 30 20 10 0 2014 2015 2016 2017 2018 CEVA Proprietary Information 2019 2020 Source: ABI Research, August 2014 3 But That’s Just Part of the Story Many possible combinations of connectivity, PER DEVICE! Google Nest Learning Thermostat Bluetooth Low Energy 802.15.4 (Thread) WiFi 802.11n Connect to your Smartphone Connect to your Home Network Connect to Internet CEVA Proprietary Information Apple Watch Bluetooth Dual Mode Connect to iPhone Wi-Fi 802.11n Speed up data transfer when needed 4 The Ultimate Connected Device… Wearalone Smartwatch Other Design Considerations Samsung Gear S2 Concurrency: Do you need to run different protocols at the same time? Different approaches: Your Smartphone is no longer necessary! WiFi 802.11n Bluetooth 4.1 1. One-time switch For example, a device that on wake-up can be configured to run either Zigbee or BLE 2. Time-slicing For example, the device is continuously switching between Zigbee and BLE, trying to avoid transmission losses 3. Simultaneous True multi-mode networking, requiring a multiple radios or an SDR approach RF integration – pros and cons GPS Cellular Do you need to futureproof your design? SW-based approach vs. lowest-cost design CEVA Proprietary Information 5 CEVA Connectivity & Communication Portfolio CEVA provides a wide range of connectivity platforms LTE-MTC LTE Cat-1, Cat-0, Cat-M LTE, Bluetooth, Wi-Fi, 802.15.4/g Complete solutions – HW, SW, PHY, MAC, reference architectures Concurrency, coexistence Available in: HW-based design, most powerefficient implementation SW-based design (SDR-like), enabling flexibility and adaptability CEVA Proprietary Information 6 How Can IoTs Get Smart(er)? Integrate (more) sensors Challenges: Data - Always-sensing requires extreme low power Signal processing for cleaning up the noise Smart Decision Making Data Fusion - Information Could require significant signal processing Memory requirements CEVA Proprietary Information - Artificial intelligence algorithms Extreme processing requirements 7 IoT Evolution: Smart Home Security From This: To This: Bluetooth 4.1 Wi-Fi 802.11n Breaking Glass Always-on Face detection CEVA Proprietary Information 8 IoT Evolution: Self Driving Cars From This: To This: And This (ADAS): 8-12 Different Image Sensors Driver Monitoring Sensor Fusion Log/Mid Range Radar Ultrasound Sensors Car-to-Car (V2V, V2X) CEVA Proprietary Information 9 IoT Evolution: Drones From This (current generation): 5 cameras!! Obstacle avoidance To This (next generation): Ultrasonic Sensors Autonomous Pilot CEVA Proprietary Information Auto Tracking Depth Mapping 10 IoT Layers – Local Intelligence vs. Cloud Local intelligence enables: Camera/microphone/other sensors raw data does not need to be sent to the cloud, only processed meta-data is being sent Increased privacy Reduced data bandwidth, transfer overhead and processing latency to/from cloud lower on-device power + lower data cost Immediate and continuous availability of local processing Quick response for latency-critical processing, no cloud availability concerns Efficient processing for scene analysis (sound/vision) with lower power than GP CPU/GPU Lower power consumption, longer battery life Local intelligence is key for smart IoT devices ! CEVA Proprietary Information 11 How Smart Can Devices Get? 1. With 10+ sensors in an average IoT, these devices are already well-aware 2. The Challenge: Recognizing and interpreting Neural Networks in embedded devices The secret ingredient: developing good training database and a lightweight neural net OCR General 2011 2016 Computers could not tell the difference between a dog and a cat Can tell the difference between different dog breeds! Could not run on-device (performance, power); had to rely on cloud servers Can run on-device by offloading the Neural Net processing onto a dedicated vision processor Various recognition and classification algorithms on the device Google Translate runs complete Neural Net algorithm on the device Machine Learning and Neural Net algorithms only in the cloud Had developed unique training database and Neural Network that fits into embedded devices CEVA Proprietary Information 12 Back to the Quiz: What’s Common for: Meta-data provided with the database was biased against gay Training database had flaws. Google removed the “gorilla” tag altogether… Very hard to meet: “Socially Acceptable Driving” CEVA Proprietary Information 13 Driverless Cars Make Mistakes Autonomous driving cars are the absolute challenge for robots The scene is very difficult and fast changing, 10s of objects to analyze, unpredictable behaviors, road conditions, weather conditions, etc’ People’s lives depend on it! Apparently they also make mistakes: Tesla releases an overnight over-the-air SW upgrade Now Tesla cars drive autonomously on freeways CEVA Proprietary Information 14 Driverless Cars Make Mistakes More autonomous cars involved in “mistakes”: Volvo had to apologize, claiming the system wasn’t scaled… CEVA Proprietary Information 15 The Real Frontier: Can Robots Adhere to “Socially Acceptable Rules”? Always obey the law, must be “polite” Driverless car would drive “like an old lady” Cannot make eye contact, cannot see or acknowledge human gestures, cannot notice “common knowledge hints” Example: Merging lanes. Driverless car would never “step on the gas” to merge in traffic Now consider this: CEVA Proprietary Information 16 So, Can Robots Get As Smart As Humans? 1. They are always-on, always alert No more worries about driver drowsiness Texting while driving? Go ahead… 2. They react more quickly and more precisely than humans That ~1.5sec human reaction time could be fatal in some cases 3. They can use sensors beyond human capabilities For example, radar and ultrasonic sensors V2V another example – reacting to things you cannot even see BUT, with Artificial Intelligence, robots can overcome humans!! CEVA Proprietary Information 17 AlphaGo Wins Lee Sedol 4:1!! GO (weiqi), unlike other board games, has too many possible options for brute-force techniques It actually requires thinking, intuition and true skills To win, AlphaGo is using Deep Learning techniques Combining two separate neural networks to narrow down his options AlphaGo continues learning and improving by continuously playing against itself CEVA Proprietary Information 18 What’s Next? Let’s Bring the Future Closer Connect more “things” Consider the vast choices you need to make Enable more senses While keeping power low for always-sensing applications Get “things” smarter, locally Use Machine Learning algorithms to take smart decisions And, don’t try to reinvent the wheel! Use systems and platforms available off-the-shelf CEVA Proprietary Information 19 CEVA’s SW Framework to Accelerate Your Neural Network Developments CEVA Proprietary Information * vs the leading GPU-based systems 20 Thank You!! Eran Briman, VP Marketing, CEVA Email: eran.briman@ceva-dsp.com