Neural map

1. A computer-implemented method for storing data associated with an agent in a multi-dimensional environment via a memory architecture, wherein the multi-dimensional environment is an environment in which the agent moves, the method comprising:

storing, in the memory architecture of a computer memory, one or more stored features at memory coordinates corresponding to environment coordinates of the multi-dimensional environment, wherein the environment coordinates of the multi-dimensional environment correspond to locations in the environment in which the agent moves, and each unique environment coordinate corresponds to, according to a coordinate normalization function, a corresponding unique memory coordinate of the memory architecture;

(a) generating, by a set of one or more processor cores, a summary of the one or more stored features stored throughout the memory architecture;

(b) retrieving, by the set of processor cores, one or more candidate features corresponding to a current position at which the agent is located in the environment from the memory coordinates of the memory architecture corresponding to the current position of the agent in the environment;

(c) determining, by the set of processor cores, whether the one or more candidate features correspond to the summary of the one or more stored features of the memory architecture; and

(d) updating, by the set of processor cores, the memory architecture with the one or more candidate features at the memory coordinates of the memory architecture that correspond to the summary of the one or more stored features of the memory architecture.

2. The computer-implemented method of claim 1 , further comprising iteratively repeating steps (a)-(d) for each new position of the agent as the agent traverses the multi-dimensional environment.

3. The computer-implemented method of claim 1 , wherein generating the summary of the one or more features stored throughout the memory architecture comprises:

passing the memory architecture through a neural network to generate a C-dimensional feature vector, wherein C is a number of features associated with the environment.

4. The computer-implemented method of claim 1 , wherein determining whether the one or more candidate features correspond to the summary of the one or more stored features of the memory architecture comprises:

computing a query vector q t according to a current state embedding s t and a feature vector r t ; and

computing a weighted average c t according to q t and each feature vector of the memory architecture.

5. The computer-implemented method of claim 1 , wherein updating the memory architecture with the one or more candidate features at the memory coordinates of the memory architecture that correspond to the summary of the one or more stored features of the memory architecture comprises:

generating a new memory architecture comprising the data stored in the memory architecture, wherein the one or more candidate features of the new memory architecture correspond to one or more stored features written to corresponding memory coordinates of the new memory architecture.

6. The computer-implemented method of claim 1 , wherein the current position of the agent corresponds to a current absolute position of the agent in the multi-dimensional environment.

7. The computer-implemented method of claim 1 , wherein the agent comprises a deep reinforcement learning agent.

8. A computer system for storing data associated with an agent in a multi-dimensional environment, wherein the multi-dimensional environment is an environment in which the agent moves, the computer system comprising:

a set of one or more processor cores;

a memory coupled to the processor cores, the memory storing:

a memory architecture comprising one or more stored features at memory coordinates corresponding to environment coordinates of the multi-dimensional environment, wherein the environment coordinates of the multi-dimensional environment correspond to locations in the multi-dimensional environment in which the agent moves, and each unique environment coordinate corresponds to, according to a coordinate normalization function, a corresponding unique memory coordinate of the memory architecture; and

instructions that, when executed by the processor, cause the computer system to:

(a) generate a summary of the one or more features stored throughout the memory architecture;

(b) retrieve one or more candidate features corresponding to a current position at which the agent is located from the memory coordinates of the memory architecture corresponding to the current position of the agent in the environment;

(c) determine whether the one or more candidate features correspond to the summary of the one or more stored features of the memory architecture; and

(d) update the memory architecture with the one or more candidate features at the memory coordinates of the memory architecture that correspond to the summary of the one or more stored features of the memory architecture.

9. The computer system of claim 8 , wherein the memory comprises a first memory storing the memory architecture and a second memory storing the instructions.

10. The computer system of claim 8 , wherein the instructions further cause the computer system to iteratively repeat steps (a)-(d) for each new position of the agent as the agent traverses the multi-dimensional environment.

11. The computer system of claim 8 , wherein the instructions cause the computer system to generate the summary of the one or more features stored throughout the memory architecture by causing the computer system to:

pass the memory architecture through a neural network to generate a C-dimensional feature vector, wherein C is a number of features associated with the environment.

12. The computer system of claim 8 , wherein the instructions cause the computer system to determine whether the one or more candidate features correspond to the summary of the one or more stored features of the memory architecture by causing the computer system to:

compute a query vector q t according to a current state embedding s t and a feature vector r t ; and

compute a weighted average c t according to q t and each feature vector of the memory architecture.

13. The computer system of claim 8 , wherein the instructions cause the computer system to update the memory architecture with the one or more candidate features at the memory coordinates of the memory architecture that correspond to the summary of the one or more stored features of the memory architecture by causing the computer system to:

generate a new memory architecture comprising the data stored in the memory architecture, wherein the one or more candidate features of the new memory architecture correspond to one or more stored features written to corresponding memory coordinates of the new memory architecture.

14. The computer system of claim 8 , wherein the current position of the agent corresponds to a current absolute position of the agent in the multi-dimensional environment.

15. The computer system of claim 8 , wherein the agent comprises a deep reinforcement learning agent.

16. A computer-implemented method for storing data associated with an agent in a multi-dimensional environment via a memory architecture, the memory architecture storing one or more features at coordinates corresponding to coordinates of the multi-dimensional environment, the method comprising:

(a) transforming, by a set of one or more processor cores, the memory architecture according to a velocity of the agent to thereby generate a transformed memory architecture;

(b) generating, by the set of processor cores, a summary of the one or more features stored throughout the transformed memory architecture;

(c) retrieving, by the set of processor cores, one or more candidate features corresponding to center coordinates of the transformed memory architecture;

(d) determining, by the set of processor cores, whether the one or more candidate features correspond to the summary of the one or more stored features of the transformed memory architecture; and

(e) updating, by the set of processor cores, the transformed memory architecture with the one or more candidate features at the center coordinates of the transformed memory architecture that correspond to the summary of the one or more stored features of the transformed memory architecture.

17. The computer-implemented method of claim 16 , further comprising iteratively repeating the steps (a)— (e) for each position of the agent as the agent traverses the multi-dimensional environment.

18. The computer-implemented method of claim 16 , wherein generating the summary of the one or more features stored throughout the transformed memory architecture comprises:

passing the transformed memory architecture through a neural network to generate a C-dimensional feature vector, wherein C is a number of features associated with the environment.

19. The computer-implemented method of claim 16 , wherein determining whether the one or more candidate features correspond to the summary of the one or more stored features of the transformed memory architecture comprises:

computing a query vector q t according to a current state embedding s t and a feature vector r t ; and

computing a weighted average c t according to q t and each feature vector in the transformed memory architecture.

20. The computer-implemented method of claim 16 , wherein updating the transformed memory architecture with the one or more candidate features at the center coordinates of the transformed memory architecture that correspond to the summary of the one or more stored features of the transformed memory architecture comprises:

generating a new memory architecture comprising the data stored in the transformed memory architecture, wherein the one or more candidate features of the new memory architecture correspond to one or more stored features written to the center coordinates of the new memory architecture.

21. The computer-implemented method of claim 16 , wherein the velocity of the agent is based on movement of the agent in the multi-dimensional environment between time steps.

22. The computer-implemented method of claim 16 , wherein the agent comprises a deep reinforcement learning agent.