Amit Mate

Founder & CEO
GMAC Intelligence, formerly at @Qualcomm Research

[Amit Mate @amitmate2010]

Excited to present at the "Embedded Vision Summit 2025" in the "Technical Insights Track"! Join me to explore the engineering challenges and solutions in building multi-modal agentic applications on low-power edge platforms, focusing on architectures for autonomy and adaptability. #AI #EdgeComputing #EmbeddedVisionSummit

[Embedded Vision Summit @EmbVisionSummit]

At the Summit, Amit Mate, CEO of GMAC Intelligence, will explore the requirements and architectures of agentic applications, including AI and non-AI requirements, and discuss two main approaches to agent-based application architecture.

https://t.co/jVHrPmjmir

[Amit Mate @amitmate2010]

@paulg @fermatslibrary Its hard to mimic intuition, but logic can be memorized and applied. If all human intelligence were exploration+exploitation, then logic is the intelligence behind exploitation and intuition is the intelligence behind exploration, necessary to make both e's very efficient.

[Amit Mate @amitmate2010]

@ericmigi @asmith my 2c - 1) touch is a major modality that needs to be integrated in the foundational models 2) another key missing piece of hardware is smart actuator - an actuator that can sense and manipulate simultaneously

[Amit Mate @amitmate2010]

@fchollet is it a practical consideration, due to earthquakes?

[Amit Mate @amitmate2010]

@keshigeyan amazing work @keshigeyan , is it possible to run the gemini 1.5 demo in colab?

[Amit Mate @amitmate2010]

https://t.co/78EkO1Gkm3

[Amit Mate @amitmate2010]

Startup is like life, you can learn from other's experiences, but you still have to live your own. My first startup memoir inspired by @paulg essays below:

[Amit Mate @amitmate2010]

@fchollet good luck with your new endeavor @fchollet and thank you for building keras which made ML accessible to the rest of us

[Amit Mate @amitmate2010]

@sanjeevsanyal very interesting phenomenon indeed. could that be a direct result of overseas venture capital fleeing china, esp. after covid? maybe its also becoming expensive for entrepreneurs to start on their own ? some more data on this would help.

[Amit Mate @amitmate2010]

Now published by @EdgeAIVision: the video of my talk at this year's @EmbVisionSummit! https://t.co/K6KO1rIu0D

[Edge AI and Vision Alliance @edgeaivision]

New Embedded Vision Summit Presentation from GMAC Intelligence: "The Fundamentals of Training AI Models for Computer Vision Applications" https://t.co/pIQaTHJhl2

[Amit Mate @amitmate2010]

@emollick Mapping a "customer order" to menu-items with customizations is *NOT* a generative task, but a discriminative task. Handling anything else is a generative task. My 2c - what is needed is the right-combination of older and newer NLP.

[Amit Mate @amitmate2010]

@fchollet and experiments are becoming expensive, so its money-bound ?

[Amit Mate @amitmate2010]

@sanjeevsanyal @shekharkapur agree and the diesel lawn-mower 😂

[Amit Mate @amitmate2010]

@MrigendraShahi @sanjeevsanyal These kind of problems can happen everywhere as similar systems are used, can be more challenging to handle in India though..

[Amit Mate @amitmate2010]

@sanjeevsanyal its probably automated gate system failure due to connectivity breakdown ? can be solved by hybrid edge+cloud for biometric access control. its cheaper/more-secure to have some functionality on premise. it dosent take a whole lot to cache a billion finger-prints or faces.

[Amit Mate @amitmate2010]

@jimkxa @__tinygrad__ AI needs both efficient integer matrix-multiplication accelerator (DSP) and floating-point parallel operation accelerator (GPU) in a single processing unit, isnt it?

[Amit Mate @amitmate2010]

Happy to share that I'll be speaking at 2024 Embedded Vision Summit! Make sure to attend it on May 21. https://t.co/a4DpNSJfmD

[Edge AI and Vision Alliance @edgeaivision]

Don't miss Amit Mate, Founder and CEO of GMAC Intelligence, at the 2024 @EmbVisionSummit. He'll cover essential aspects of training convolutional neural networks, including prerequisites and the model training process.

https://t.co/tdy91zDpRK

[Amit Mate @amitmate2010]

@rao2z It's just the nature of error based learning systems?
https://t.co/2DZSWcJPyJ

[Amit Mate @amitmate2010]

@ylecun a system that learns from data (with errors), where learning itself is based on error propagation, will be prone to hallucinations: mathematically P(Y/X)=P(X/Y)*P(Y)/P(X) biases output towards a certain Y for very small P(X), assuming training data had spurious associations??

[Amit Mate @amitmate2010]

@rao2z is hallucination limited to language models ? If not, the explanation has to be more generic ?