Suggestions
Faisal Mohammad
Machine Learning Engineer at The Johns Hopkins University Applied Physics Laboratory
Faisal Mohammad is a Senior Machine Learning Engineer at the Johns Hopkins University Applied Physics Laboratory (APL). He specializes in machine learning and computer vision, contributing to various projects within the laboratory. His professional background includes experience in developing models and algorithms for applications in fields such as energy load forecasting using deep neural networks.
In addition to his role at APL, Faisal has a presence on platforms like LinkedIn, where he shares insights related to his work and expertise in machine learning technologies. His contributions to the field are also reflected in academic publications, indicating an active engagement in research and development initiatives.
For more detailed information about his professional journey, you can view his LinkedIn profile here .1
Highlights
Is cluttering your Linkedin profile with the keyword 'LLM' the next move? 🤔
Imagine this highly simplified RL experiment:
given an n-order polynomial, we train an actor-critic model to differentiate it. Each iteration, the polynomial is parsed and the model performs a series of silly math operations as its actions.
For example if we want to differentiate x^2. We provide several options:
- leave it as is
- move the exponent in front of the x
- copy the exponent in front of the x and add 1 to it
- set everything to 0
- add 1 to the exponent and do nothing else
- blah blah blah
- ...
and we penalize it or reward it for differentiating it correctly. I'd imagine it would eventually learn to produce the differentiation properly. Now does it work? Yeah it will always work given how simple it is. But grouping it in the category of reasoning, comprehension, and understanding is completely far-fetched. It has no idea what differentiation is. We've simply, by brute force, made it seemingly learn the power rule for differentiating polynomials.
So in extension, slapping together word token representations and visual token representations and telling a robot where to navigate is not enabling reasoning in visual language navigation.
Feeding ground truth data of question and answers and training a model to select the correct reason by brute force (backprop) should not be called Visual Commonsense Reasoning.
If we're attaching words like 'reasoning' and 'comprehension' in ML/AI conference papers, we ought to be explicit/honest with the way it is worded.
An honest statement would be "we've designed a model that's seemingly able to exhibit that it is reasoning".
ok rant over 😀😀😀
