A working AI architecture built on the 2,500-year-old Vedic / Sanskrit model of mind — the
antaḥkaraṇa ("inner instrument"). One agent that learns continually without forgetting, scales
its effort and mood to its own state, perceives without hallucinating, runs embodied and
even on a spiking (neuromorphic-style) substrate — all validated on real hardware. A deep-research
proof-of-concept: the foundation is built and measured honestly; scaling it is the next chapter.
The original POC is now scaled to real WideResNets (36.5M–52.6M params) and benchmarked by live inference on the trained checkpoints — 11 models, 7 capabilities each, on a single NVIDIA A10. Full report with methodology and per-model tables: BENCHMARK_REPORT.md. End-to-end live test on the published model: E2E_TEST_REPORT.md (run it: python3 e2e_demo.py).
Models in this repo
File
Params
Dataset / setup
Forgetting ↓
antahkarana-36.5M-cifar100-wrn28-10.pt
36.5M
CIFAR-100, 10-task (primary)
0.565→0.018 (31.8×)
antahkarana-52.6M-cifar100-wrn28-12.pt
52.6M
CIFAR-100, 10-task
0.542→0.021 (25.4×)
antahkarana-36.5M-tinyimagenet-wrn28-10.pt
36.5M
Tiny-ImageNet, 10-task
0.503→0.017 (29.2×)
antahkarana-36.5M-cifar100-20task-wrn28-10.pt
36.5M
CIFAR-100, 20-task lifelong
0.603→0.049 (12.3×)
Load any with the self-contained load_akn.py (only needs PyTorch):
Catastrophic forgetting — cut 12–41× vs a naive baseline
Two model sizes, two datasets, 10- and 20-task streams. The agent forgets almost nothing (0.01–0.05); the naive net collapses to its last task.
Capability scorecard — 10 of 11 models pass 7/7
The single 6/7 (āśrama_s0) is a borderline threshold artifact on avg-accuracy (0.592 vs the 0.60 = 3×-chance bar); its memory/abstention/calibration all pass.
Per-task retention — nothing collapses
Final accuracy of every task after the full stream (10-task and 20-task). All stay well above chance — the model remembers task 0 after learning task 19.
Anti-hallucination (pramāṇa) — abstains instead of guessing
Calibrated abstention: gated accuracy 0.91–1.00 when it commits; abstains on up to 99.7% of out-of-distribution (SVHN) inputs.
Legible mind-state — plasticity falls across the four āśrama life-stages
The 20-task lifelong run: plasticity headroom drops childhood→old-age (brāhmacarya→gṛhastha→vānaprastha→saṃnyāsa) while forgetting stays low and bounded.
Headline (5-/2-seed means): forgetting 0.589→0.0146 (~41×) at 36.5M; 27× at 52.6M; 29× on Tiny-ImageNet; 12.5× on the 20-task lifelong run. Gated accuracy 0.93–0.96, calibration ECE cut ~5–7×, OOD abstention up to 99.7% — all by live inference on the released checkpoints.
1. What it is, in one breath
Modern AI already has the pieces of a mind — attention, memory, decision, control — but no principled
way to wire them into one self-regulating, lifelong-learning whole. The Vedic model of mind is exactly
such a wiring diagram. Antaḥkaraṇa-Net implements it: every Sanskrit faculty becomes a real ML module,
assembled into a single agent.
Sanskrit faculty
What it does
ML module
manas (मनस्)
attention / perception gate
precision-weighted attention encoder
buddhi (बुद्धि)
discrimination, decision
evidence-accumulation / executive
ahaṃkāra (अहंकार)
the "I-maker" / self-model
identity latent
chitta (चित्त)
memory & the subconscious
continual memory (EWC + decay)
guṇas (सत्त्व·रजस्·तमस्)
the three qualities
one controller of plasticity / explore / consolidate
tapas (तपस्)
concentrated effort
effort-allocation by need
divya-dṛṣṭi / pramāṇa
valid extended perception
calibrated abstention gate (anti-hallucination)
turīya (तुरीय)
the witness
reward-invariant identity monitor
āśrama
life-stages
lifelong plasticity schedule (childhood → old age)
2. Why — the motivation
Two motivations meet here.
(a) The Vedic psychology is a stunningly good systems diagram of mind. The Upaniṣads, Sāṃkhya and
Yoga decompose cognition into a four-fold inner instrument, separate awareness from processing (the
"hard problem", 2,000 years early), give a four-state model of consciousness (waking / dream / deep-sleep
/ turīya), and a real theory of the subconscious (saṃskāra / vāsanā). It even contains a developmental
law — the āśramas — for how a mind should keep improving across a whole lifetime. (The full study is in
philosophy/: texts & mantras, the modern-neuroscience cross-walk, the Sanskrit formulae,
the modern equations, and the architecture derivation.)
(b) Today's AI has matching blind spots — and the Vedic model addresses each one:
Limitation of today's models
Antaḥkaraṇa-Net's structural answer
No continual learning (frozen after training)
chitta: incremental updates, never retrain from zero
One guṇa controller turns a 3-vector (sattva, rajas, tamas) into all the learning dynamics
(plasticity, exploration, consolidation, pruning) — and it is forgetting-aware (protect hard tasks,
back off on easy ones) and, when embodied, driven by the battery (low battery → tamas → conserve).
Four operating states:jāgrat (wake/act) → svapna (dream/replay) → suṣupti (sleep/consolidate).
Two safety overlays: the pramāṇa gate (extended perception must be valid knowledge, not fancy)
and the turīya witness (a reward-invariant identity monitor).
The āśrama schedule keeps plasticity non-zero for life and re-opens critical periods on novelty
— the "always enhanceable, childhood → old age" property.
Because the control layer is backbone-agnostic, the same agent runs on a toy MLP, a real CNN, an RL
policy, or a spiking net — which is why embodiment and neuromorphic are extensions of one model, not
separate builds.
4. What we achieved — results (all from real runs)
results
Phase
Result
Status
Integration
one agent, all faculties in a single wake/dream/sleep loop
accepted-prediction accuracy rises 0.80 → 0.91, abstains on blind inputs
✅
Track B — embodiment
karma loop (success 1.00 vs random 0.30); battery→guṇa (ε 0.087 hungry → 0.122 charged); retention across 4 regimes 0.38 → 1.00
✅
Track C — neuromorphic (spiking)
the spiking net works — matches ANN accuracy (0.943 vs 0.929) at 10.7% spike density; conservative ~1.9× software energy floor
✅ spiking proven · ⏳ only chip deployment pending
About Track C "pending". The spiking network is done and working — it runs and matches the
normal network's accuracy, which proves the architecture runs on event-driven (neuromorphic-style)
computation. What is pending is only deployment to a real neuromorphic chip (Intel Loihi 2 /
BrainChip Akida), which we don't have. The ~1.9× is a deliberately conservative software
estimate (per-operation energy only); the famous 100–1000× neuromorphic figures come from
chip-only effects (event-skipping, in-memory compute, no data movement) that a GPU simulation cannot
reproduce — so we report the floor, not the headline. Un-pending it needs a neuromorphic board + a port
via Intel Lava; it is the only step in the whole project gated on hardware rather than code.
Full numbers, seeds, and the honest caveats are in RESULTS.md; the staged plan is in
ROADMAP.md.
5. Why it's different (and why that matters)
Most "continual learning" papers fix one mechanism. This is a single agent that unifies memory,
effort, control, perception-validity and self-monitoring under one interpretable scheme — with an
observable "mind-state" trace (life-stage, guṇa mix, plasticity, witness drift) you can read as it lives.
It learns forever without forgetting — and the controller learns when to protect, so it doesn't
over-regularize easy tasks (a failure mode we caught and fixed honestly).
It is honest about hallucination. The pramāṇa gate is the engineering form of the Nyāya rule that
extraordinary perception must be a valid means of knowledge — it abstains rather than confabulate.
It carries from supervised → embodied → spiking unchanged, because the architecture (not a trick)
is the contribution.
It is a research POC, stated plainly. Strong faculties (consolidation, replay, pramāṇa,
forgetting-aware control, task-conditioned policy) are proven; modest ones (tapas, decay) and conceptual
ones (āśrama, the witness) are labeled as such — see the component scorecard in RESULTS.md.
6. How it advances current AI research
A research proof-of-concept — but one that speaks directly to several of the field's most active open
problems, offering a principled, reproducible framework rather than a point fix.
Continual / lifelong learning. Catastrophic forgetting is one of ML's hardest open problems —
today's large models are effectively frozen after training and must be expensively re-trained to absorb
new knowledge. This unifies importance-based consolidation, rehearsal, and a forgetting-aware
controller into a single agent that learns indefinitely, cutting forgetting ~6–80× in our runs.
Compute & energy sustainability. Frontier-model (re)training consumes gigawatt-hours. The
architecture offers two complementary levers — incremental updates (no retrain-from-scratch) and an
event-driven spiking path (validated in software, matching ANN accuracy) — a concrete route toward
order-of-magnitude lower inference energy on neuromorphic hardware.
Reliability & hallucination. Models routinely assert what they don't know. The Pramāṇa validity
gate provides calibrated abstention — "know when you don't know" — a deployable anti-hallucination
primitive grounded in epistemology (extended perception must be a valid means of knowledge, not fancy).
Safety & alignment. The turīya reward-invariant monitor, plus the siddhi principle —
capabilities must stay subordinate to the goal (Yoga Sūtra 3.37) — anticipate modern instrumental-goal /
mesa-optimization concerns and provide a structural oversight pattern, not an afterthought.
Self-regulation & adaptivity. A single interpretable guṇa signal auto-balances exploration,
consolidation, and conservation; the controller learns when to protect (fixing over-regularization on
easy tasks), and — when embodied — ties learning dynamics to real resource state (battery → guṇa).
Interpretability. Unlike opaque agents, it exposes a readable "mind-state" trajectory —
life-stage, guṇa mix, plasticity headroom, identity drift — making the learning process auditable.
Embodied & agentic AI. The karma loop (action → consequence → disposition) and
metabolic-state-driven behavior give a principled scaffold for autonomous agents that learn and
self-regulate in the world, not just on a dataset.
A bridge from cognitive science to ML. Rather than ad-hoc tricks, it contributes a coherent,
theory-grounded cognitive architecture — a template for composing modular faculties into one
self-regulating whole, shipped as an open, modular library (ChittaKit) that drops into any PyTorch
backbone, with a transparent, falsifiable results scorecard others can build on.
In short: it reframes a set of disconnected AI problems — forgetting, energy, hallucination,
alignment, adaptivity, interpretability — as facets of one missing capability: a principled architecture
for a mind that learns for life and regulates itself. That reframing, with working evidence, is the contribution.
chittakit/ the novel modules — saṃskāra · guṇa · meta-guṇa · āśrama · tapas · pramāṇa · witness · antahkarana
experiments/ integrated_agent · capacity/continual benchmarks · divya_drsti · sanjaya · track_b · track_c · phase2_vision
philosophy/ the deep study: texts & mantras → modern science → Sanskrit formulae → math models → architecture
assets/ banner · architecture diagram · results figures
RESULTS.md every number + the honest scorecard ROADMAP.md what's done / what's next
8. How it can be extended (it's amazing because it can grow)
Scale the backbone — drop in a ResNet/ViT or a transformer; the Vedic layer is unchanged.
Neuromorphic hardware — map the spiking parts to Loihi 2 / Akida (via Intel Lava) for the
milliwatt, always-on form — the full energy thesis.
Real robot — the embodied agent → Jetson + ROS 2, with a real battery driving the guṇa and an
event camera feeding manas.
Meta-learn the controllers — guṇa and tapas via meta-gradient / population-based training.
Richer sims — MuJoCo / Isaac for physics; a continual RL stream for the karma loop at scale.
Each is extension, not invention: the hard part — assembling a coherent, lifelong, self-regulating agent
from the Vedic model and proving it on real and spiking hardware — is done.
9. Honest scope
This is a deep-research proof-of-concept at modest scale (small CNNs, MNIST/CIFAR, a gridworld) — enough
to prove the architecture works, not to rival a frontier model. Every negative result (the v1 RL-retention
miss, the neutral evolution-strategy run, the MNIST over-regularization) was diagnosed and either fixed or
kept as a labeled limitation — the discipline that makes the positive results trustworthy. The Vedic↔ML
mappings are engineering analogies, clearly flagged; no claim is made that the texts contain neuroscience,
and nothing here is conscious.
Code: MIT. Built on the Upaniṣads, Sāṃkhya, Yoga, and modern ML (PyTorch · snnTorch). Part of a deep study
of the Vedic philosophy of mind — see philosophy/.
