Author: Benjamin Obi
Date: May 27, 2024
Categories: [Artificial Intelligence, Projects, Python]
There’s no shortage of AI chatbots today. Most of them are brilliant, powered by massive Large Language Models (LLMs) that have digested a significant portion of the internet.
But what if you want to build an AI that is efficient, transparent, and brain-inspired without relying on a 175-billion-parameter model? What if you’re going to understand and engineer intelligence from first principles?
That was the challenge I set for myself. The result is ACEP (Artificially Cognitive Educational Partner), a psychology tutor powered not by an LLM, but by a custom-built Hyperdimensional Computing (HDC) engine.
In this post, I’ll pull back the curtain on how it works.
While LLMs are incredible, they can be seen as “black boxes.” It’s often difficult to trace why they generate a specific answer. I was fascinated by an alternative paradigm called Hyperdimensional Computing (HDC).
Also known as Vector Symbolic Architectures, HDC is a computational framework inspired by the fact that the human brain computes with high-dimensional, distributed representations. It’s a model of cognition that is:
Transparent: Every operation is mathematically defined and interpretable.
Robust: Naturally resistant to noise and imperfections.
Efficient: Operations are simple and highly parallelizable.
I wanted to see if I could use this paradigm to build a functional, focused AI application. A psychology Q&A tutor seemed like the perfect use case.
So, how do you give a computer a brain-like associative memory? Here’s the high-level architecture of ACEP:
The first step was to create a knowledge base. I curated a comprehensive dataset of psychology concepts from classical conditioning and cognitive dissonance to memory formation and social learning theory.
In HDC, concepts aren’t stored as text; they’re represented as hypervectors, long, random vectors (e.g., 10,000 dimensions) with values of -1 and 1. The key insight is that these random vectors are nearly orthogonal, meaning each one is unique and can be combined and retrieved without confusion.
The concept "classical conditioning" gets its own unique hypervector.
So does "Pavlov", "bell", and "salivation".
Simple concepts are combined to form complex ideas using three core HDC operations:
Bundling (+): Represents a set of items. The hypervector for "learning" might be a bundle of "conditioning", "reinforcement", and "observation" vectors.
Binding (*): Represents the association between items. The relationship "Pavlov discovered classical conditioning" is formed by binding the vectors for Pavlov and classical conditioning.
Permutation: Represents sequence and order, crucial for encoding timelines or steps in a process.
These operations allow the system to build a rich, interconnected web of psychological knowledge stored in a high-dimensional space (vector_store.py).
This is where the magic happens. When a user asks a question like “How does stress affect memory?”:
The query processor breaks down the question into its core concepts: ['stress', 'affect', 'memory'].
Each concept is mapped to its corresponding hypervector.
These vectors are combined to form a query hypervector that represents the meaning of the question.
The system performs a similarity search across the entire memory store to find the hypervector(s) that most closely match the query.
The reasoning engine then takes these matched concepts and constructs a coherent, natural language response, complete with a confidence score.
The entire process is a dance of geometry in a 10,000-dimensional space, all happening in milliseconds.
The system is built as a modular Python framework:
hdc_core.py: The beating heart. A clean implementation of the HDC algebra.
vector_store.py: The long-term memory, handling storage and retrieval of hypervectors.
reasoning_engine.py: The prefrontal cortex. It classifies query intent (is this asking for a definition, a comparison, or a cause?) and formulates the response.
main.py: The friendly face is a sleek Streamlit web app that provides the UI.
The final product is a responsive web application that provides accurate, insightful answers to psychology questions. It can explain complex theories, compare concepts, and provide examples, all with a measured confidence score for its answers.
It’s not just a simple chatbot; it’s a purpose-built knowledge retrieval and reasoning system.
Building ACEP was more than a coding exercise; it was a deep dive into an alternative future for AI.
Understandability is a Feature: Being able to trace the system’s “thought process” from query to answer is incredibly powerful and is a key advantage of HDC over more opaque models.
Efficiency on the Edge: This architecture is incredibly lightweight. It hints at a future where sophisticated AI can run on low-power devices without needing a constant internet connection to a cloud API.
The Principles are Key: Working with HDC forces you to think rigorously about how knowledge is represented and manipulated a fundamental skill for any AI engineer.
This project solidified my understanding of core AI principles and demonstrated my ability to architect and implement a complex, end-to-end AI system from scratch.
ACEP is a living project, and I’m excited to share it.
👉 Live Demo: Ask it a psychology question!
👨💻 Source Code on GitHub: Dive into the implementation details. Contributions and ideas are welcome.
This is just the beginning. The potential applications for HDC in robotics, IoT, and other areas where efficiency and explainability are critical are vast. I’m excited to continue exploring this fascinating intersection of neuroscience, computer science, and software engineering.
