4 ms·

Grid cells is, according a learned colleague of mine, an epiphenomenon. If so, this is not a viable path, as it is really just the grandmother cell theory in di

Grid cells is, according a learned colleague of mine, an epiphenomenon. If so, this is not a viable path, as it is really just the grandmother cell theory in disguise. I like Hawkin's work. I think he has good instincts, and he was probably the first to hammer home the point that all the brain does is make predictions and there are a small number of evolutionary mechanisms at work in the brain. Still, that doesn't make him right here.

clickok 8y ago
I am not sure that he's the first to note that the brain is a predictive machine, or to notice that models of state or the environment can be formulated entirely in terms of predictions[0]. Personally, I find that idea very convincing, because we already know how to learn to predict (reinforcement learning), and we can use the same framework to optimize an agent's behavior to maximize against these predictions. There is some evidence that the brain uses something like temporal difference methods to learn. The hard parts are: how to do this in a stable and efficient matter, and how to build up our representation to go from relatively simple goals (e.g., survive and collect resources) to more abstract goals (e.g., translate a poem in such a way that the original's charm is preserved[1]). That is, you want to have a system that is capable of climbing up Maslow's hierarchy (or something like it) and actually proceeds to do so based on emergent properties of the system itself. Proposing and investigating new ways of constructing learning systems so that there is a continuum between primal drives and complex abstract goals is hard, and even if grid cells are an epiphenomena, then it is worth studying them to see how they might develop from simpler rules. If Hawkins' team can make something useful, then that can still be beneficial by allowing for more powerful and flexible AI, even if it's not a theory of everything. --- 0. See for example: http://web.eecs.umich.edu/~baveja/Papers/psr.pdf http://web.eecs.umich.edu/~baveja/Papers/psr.pdf (Littman, Sutton, and Singh), which describes and analyzes the equivalence between belief-state based representations and state representation based on predictions.
- jamesblonde 8y ago
  You have a point that the grandmother cell theory appeared first. Got people thinking. Then the grandmother cell theory people themselves came out and said 'hey, we were wrong - we think it is sparse coding' https://en.wikipedia.org/wiki/Alternative_explanations_of_the_%22grandmother%22_cell https://en.wikipedia.org/wiki/Alternative_explanations_of_th... Something similar could happen here. But then, it won't be grid cells as they are. Will it be that location is coded in every sparse representation? I don't know...
diminish 8y ago
Could you elaborate why an epiphenomenon isn't viable? Being based on grandmother cell theory seems like an advantage.
- jamesblonde 8y ago
  In the grandmother cell theory people said - hey this cell only seems to fire when i show you a picture of your grandmother. Hence all the hierarchical features that make up your mental representation of your granny are found in the extraordinary binding properties of this one neuron. Of course, if you know large sparse DNNs, you know that you have 50 input vectors and pick any neuron far from the input, it may only activate in response to one of the 50 input vectors. It doesn't mean that it is the input vector. Similarly, Jeff has a theory that all representations in the brain include spatial representations. Now he claims the grid cell does that. But how would the grid cell have evolved to do provide the spatial coordinates? Could not the spatial coordinates be part of the sparse coding for the entity in question - why do they need the grid cells?
  - lucidrains 8y ago
    I think you misunderstood the paper. Grid cells don't encode location, a set of grid cell modules do. So there isn't a grandmother cell here at all. He's simply proposing this theory because of the noted presence of grid cells in the neocortex. His new theory, if I understand it correctly, is pretty basic. It is simply saying every sparse encoding of sensory inputs is also tied to a spatial encoding by sets of grid cell modules within the same cortical column. For deep learning practitioners, it would be like saying every embedding comes with it's own spatial encoding, or placed into the same embedding space, however you want to look at it. These spatial encodings are then related to other spatial encodings of other sensory inputs on different cortical column stacks (by displacement cells?), to firstly build up representations of objects, and then, he surmises further, the relationships of these objects with each other (part to whole). I think the last part is probably what Jeff will need to substantiate a bit more.
    - monocasa 8y ago

clickok 8y ago

I am not sure that he's the first to note that the brain is a predictive machine, or to notice that models of state or the environment can be formulated entirely in terms of predictions[0]. Personally, I find that idea very convincing, because we already know how to learn to predict (reinforcement learning), and we can use the same framework to optimize an agent's behavior to maximize against these predictions. There is some evidence that the brain uses something like temporal difference methods to learn. The hard parts are: how to do this in a stable and efficient matter, and how to build up our representation to go from relatively simple goals (e.g., survive and collect resources) to more abstract goals (e.g., translate a poem in such a way that the original's charm is preserved[1]). That is, you want to have a system that is capable of climbing up Maslow's hierarchy (or something like it) and actually proceeds to do so based on emergent properties of the system itself. Proposing and investigating new ways of constructing learning systems so that there is a continuum between primal drives and complex abstract goals is hard, and even if grid cells are an epiphenomena, then it is worth studying them to see how they might develop from simpler rules. If Hawkins' team can make something useful, then that can still be beneficial by allowing for more powerful and flexible AI, even if it's not a theory of everything. --- 0. See for example: http://web.eecs.umich.edu/~baveja/Papers/psr.pdf http://web.eecs.umich.edu/~baveja/Papers/psr.pdf (Littman, Sutton, and Singh), which describes and analyzes the equivalence between belief-state based representations and state representation based on predictions.

jamesblonde 8y ago
You have a point that the grandmother cell theory appeared first. Got people thinking. Then the grandmother cell theory people themselves came out and said 'hey, we were wrong - we think it is sparse coding' https://en.wikipedia.org/wiki/Alternative_explanations_of_the_%22grandmother%22_cell https://en.wikipedia.org/wiki/Alternative_explanations_of_th... Something similar could happen here. But then, it won't be grid cells as they are. Will it be that location is coded in every sparse representation? I don't know...

diminish 8y ago

Could you elaborate why an epiphenomenon isn't viable? Being based on grandmother cell theory seems like an advantage.

jamesblonde 8y ago
In the grandmother cell theory people said - hey this cell only seems to fire when i show you a picture of your grandmother. Hence all the hierarchical features that make up your mental representation of your granny are found in the extraordinary binding properties of this one neuron. Of course, if you know large sparse DNNs, you know that you have 50 input vectors and pick any neuron far from the input, it may only activate in response to one of the 50 input vectors. It doesn't mean that it is the input vector. Similarly, Jeff has a theory that all representations in the brain include spatial representations. Now he claims the grid cell does that. But how would the grid cell have evolved to do provide the spatial coordinates? Could not the spatial coordinates be part of the sparse coding for the entity in question - why do they need the grid cells?
- lucidrains 8y ago
  I think you misunderstood the paper. Grid cells don't encode location, a set of grid cell modules do. So there isn't a grandmother cell here at all. He's simply proposing this theory because of the noted presence of grid cells in the neocortex. His new theory, if I understand it correctly, is pretty basic. It is simply saying every sparse encoding of sensory inputs is also tied to a spatial encoding by sets of grid cell modules within the same cortical column. For deep learning practitioners, it would be like saying every embedding comes with it's own spatial encoding, or placed into the same embedding space, however you want to look at it. These spatial encodings are then related to other spatial encodings of other sensory inputs on different cortical column stacks (by displacement cells?), to firstly build up representations of objects, and then, he surmises further, the relationships of these objects with each other (part to whole). I think the last part is probably what Jeff will need to substantiate a bit more.
  - monocasa 8y ago