TLDR: This paper details the integration of Holographic Declarative Memory (HDM), a vector-symbolic memory system, with Lisp ACT-R, a widely used cognitive architecture. The adaptation allows ACT-R models to leverage HDM’s benefits like scalability and architecturally defined similarity, including a novel method for full-chunk memory retrieval using vector representations and temporal encodings, without major changes to existing models. This paves the way for more complex cognitive agents, especially for modeling sociocultural influences on decision-making.
Researchers Meera Ray and Christopher L. Dancy have introduced a significant advancement in cognitive modeling by successfully integrating Holographic Declarative Memory (HDM) with Lisp ACT-R, a widely used cognitive architecture. This work bridges a newer vector-symbolic memory model with decades of cognitive modeling, aiming to enhance the capabilities of cognitive agents, particularly for simulating complex human behaviors like disaster survivor decision-making.
Understanding the Core Technologies
ACT-R (Adaptive Control of Thought—Rational) is a cognitive architecture designed to simulate human cognition. It features two primary memory components: procedural memory (rules for actions) and declarative memory (facts and experiences). Traditional ACT-R declarative memory (DM) has limitations, such as a lack of partial memory recall (it’s either a perfect match or nothing) and scalability issues when dealing with large amounts of information, like entire text corpora.
Holographic Declarative Memory (HDM) offers a vector-symbolic alternative to ACT-R’s DM. Instead of storing entire chunks of information, HDM represents each unique input token in a vector space. This approach provides several advantages: continuous similarity measures between queries and memories, improved scalability for large datasets, and architecturally defined similarity between memory chunks. Essentially, HDM allows a cognitive model to “read” and process vast amounts of text, making it a distributed semantics model akin to Word2vec, but integrated as a memory system within a cognitive architecture.
Why Integrate HDM with ACT-R?
The integration of HDM addresses key limitations of ACT-R’s default memory system. For models that need to represent complex sociocultural structures or implement theories relying on context-dependent associations, such as Instance-Based Learning (IBL) Theory, HDM’s vector-based representation is highly beneficial. It also simplifies potential future integrations with generative models, which commonly use vector-based representations for text.
The researchers’ long-term goal is to build a cognitive agent that can simulate the actions and decisions of disaster survivors, taking into account social factors and situational awareness. HDM’s ability to process and represent large textual contexts makes it ideal for incorporating background knowledge and worldviews that influence decision-making.
Key Adaptations and Innovations
To achieve this integration, several adaptations were necessary. The team used ACT-R 7 Python connection files and existing HDM Python code to implement Lisp-side ACT-R commands. Core commands for adding and retrieving memory were implemented, and the `dm` command was updated to visualize the HDM vector space rather than explicitly stored chunks. For precise recall of goal chunks, ACT-R’s default `define-chunk` mechanism is still used.
A significant part of the work involved creating a text processing pipeline. A new ACT-R command, `preprocess-text`, uses the Natural Language Toolkit (NLTK) to remove stopwords and tokenize text into sentences. This preprocessed text can then be read into HDM using the `read-corpus-hdm` command, allowing models to ingest large documents efficiently.
Perhaps the most novel contribution is the development of a mechanism for **full-chunk retrieval** from HDM. Unlike ACT-R’s DM, HDM originally lacked a straightforward way to retrieve an entire chunk of memory based on a partial cue without knowing every slot. The researchers devised a method using neural oscillator encodings, a biologically plausible theory of temporal encoding, to represent when bits of memory were entered together. These time encodings are bound with HDM’s memory vectors using a fractional binding operation, creating `time-memory (mt)` vectors. When a cue is provided, a reconstructed time vector is generated, which is then compared against the model’s continuous time function to identify and retrieve the entire associated chunk.
Also Read:
- Unlocking Deeper LLM Understanding: A New Approach to Active Memory Management
- Boosting Active Inference Performance on Hardware for Edge AI
Preliminary Findings and Future Directions
Preliminary results indicate that the time vector representations, while noisy, provide distinct representations for different chunks. The similarity function between time vectors shows a gradual decrease as time steps grow apart, along with oscillations that represent cognitive error in serial memory recall, consistent with human memory characteristics. Further research is planned to quantify this noise and optimize parameters for better fitting observed cognitive effects.
This work maintains the vector-symbolic advantages of HDM, such as chunk recall without storing the actual chunk and improved scaling, while extending it to allow existing ACT-R models to work with the system with minimal modifications. The researchers plan to explore alternative time encodings, examine different chunk retrieval mechanisms (potentially using clustering measures), and further integrate generative models within ACT-R. This integration holds profound implications for expanding sociocultural representations in cognitive architectures, enabling more nuanced modeling of how social structures influence human decision-making. For more details, you can read the full paper here.
