Datasets:

mhough
/

inceptionv1-tuning-atlas

layer stringclasses 12 values	unit int64 0 1.02k	osi float64 0 1	pref_orientation_deg int64 0 165	pref_curvature float64 0 3	curve_selectivity_index float64 -1 1	response_magnitude float64 0 7.45	null_response_magnitude float64 0 0.14	null_alive bool 2 classes	null_osi_mean float64 0 0.84	null_osi_std float64 0 0.54	null_osi_max float64 0 1	osi_z_vs_null float64 -6.07 8.77k	osi_exceeds_all_nulls bool 2 classes	shuffle_osi_mean float64 0 0.89	shuffle_osi_max float64 0 1	osi_exceeds_all_shuffles bool 2 classes	shuffle_alive bool 2 classes	pref_freq_railed null	null_csi_mean float64 -0.21 0.47	csi_z_vs_null float64 -15,196.52 256k
conv1	0	0.299991	45	2	0.028395	3.298291	0.097651	true	0.080818	0.068984	0.192103	3.177149	true	0.050934	0.081527	true	true	null	-0.000341	15.630551
conv1	1	0.032887	0	2	0.118545	0.227065	0.142921	true	0.108162	0.092506	0.254096	-0.813727	false	0.060889	0.160188	false	true	null	0.015423	3.532939
conv1	2	0.2049	150	2	0.02486	2.442224	0.093767	true	0.140164	0.083613	0.240892	0.77423	false	0.009924	0.033787	true	true	null	-0.000967	42.406425
conv1	3	0.380811	105	0	-0.000867	1.50466	0.105561	true	0.201806	0.124198	0.414704	1.441279	false	0.094533	0.21019	true	true	null	0.009885	-0.636963
conv1	4	0.0194	120	2	0.008883	0.397106	0.092618	true	0.143775	0.074214	0.209268	-1.675905	false	0.187241	0.240801	false	true	null	0.000481	3.93033
conv1	5	0	45	2	0.053279	1.312667	0.119938	true	0.114936	0.115362	0.300449	-0.996313	false	0.093207	0.180337	false	true	null	0.043894	0.198587
conv1	6	0.367463	45	0	-0.000155	3.588955	0.117446	true	0.16158	0.067382	0.254206	3.055439	true	0.081984	0.127486	true	true	null	0.026766	-0.780065
conv1	7	0.301781	165	0	-0.000189	3.15412	0.107911	true	0.176894	0.057712	0.23734	2.163968	true	0.052699	0.093894	true	true	null	0.024529	-0.637617
conv1	8	0.30132	60	0	-0.001857	0.233685	0.109892	true	0.025419	0.007847	0.036274	35.160542	true	0.393561	0.615204	false	true	null	0.024018	-0.528972
conv1	9	0.27386	120	0.25	0.000212	3.458575	0.102165	true	0.190449	0.152341	0.457114	0.547524	false	0.065859	0.124446	true	true	null	0.03258	-0.530956
conv1	10	0.25668	135	2	0.017481	2.999887	0.109032	true	0.083685	0.041792	0.143061	4.139463	true	0.028195	0.062849	true	true	null	0.01344	0.217636
conv1	11	0.234756	45	0.25	0.004055	2.625478	0.096574	true	0.142914	0.122344	0.340201	0.750692	false	0.012177	0.020525	true	true	null	0.019352	-0.353498
conv1	12	0.396177	90	1.5	0.008892	3.381306	0.09858	true	0.164332	0.10746	0.287864	2.157494	true	0.109082	0.141045	true	true	null	0.001272	1.675044
conv1	13	0.000495	15	0	-0.000072	0.284555	0.107689	true	0.173558	0.152628	0.429073	-1.133892	false	0.150803	0.288769	false	true	null	0.040088	-0.845279
conv1	14	0.320628	135	2	0.021407	3.341014	0.096279	true	0.134628	0.094477	0.264914	1.96873	true	0.048695	0.062229	true	true	null	0.002619	4.15041
conv1	15	0.284363	105	0	-0.000888	0.047165	0.111768	true	0.146866	0.119184	0.322887	1.153652	false	0.610227	0.767284	false	true	null	0.014215	-0.431211
conv1	16	0	0	2	0.004499	0.849887	0.122667	true	0.076958	0.076747	0.201092	-1.002749	false	0.082479	0.178072	false	true	null	0.016286	-0.515226
conv1	17	0.157674	150	0	-0.000723	2.945971	0.12998	true	0.095444	0.081441	0.196099	0.764107	false	0.004653	0.009044	true	true	null	0.040881	-0.722055
conv1	18	0.315718	120	0	-0.000211	3.262272	0.097564	true	0.164505	0.135193	0.322073	1.118499	false	0.08238	0.149308	true	true	null	0.020292	-0.690027
conv1	19	0.344325	75	0	-0.000122	3.865971	0.093713	true	0.148536	0.130577	0.329315	1.499416	true	0.091627	0.213202	true	true	null	0.028376	-0.626202
conv1	20	0.362946	165	0.25	0.00071	1.632828	0.095658	true	0.239022	0.070361	0.331377	1.761272	true	0.147352	0.215917	true	true	null	0.00328	-0.37932
conv1	21	0.203805	30	0	-0.000991	2.432997	0.087405	true	0.086912	0.087387	0.235226	1.337656	false	0.033637	0.061861	true	true	null	0.003043	-0.896769
conv1	22	0.305137	150	0	-0.001764	0.225641	0.098495	true	0.162181	0.10171	0.269442	1.40552	true	0.100523	0.17093	true	true	null	0.011963	-0.98849
conv1	23	0.248464	0	0.25	0.000122	2.649346	0.09275	true	0.175705	0.146582	0.434002	0.496369	false	0.043611	0.062757	true	true	null	0.002033	-0.492802
conv1	24	0.132617	0	2	0.009035	3.108439	0.084355	true	0.130563	0.082422	0.24708	0.02493	false	0.002785	0.005173	true	true	null	0.008932	0.006988
conv1	25	0	0	0	-0.000255	1.56148	0.13547	true	0.073503	0.07088	0.190103	-1.037008	false	0.11862	0.185947	false	true	null	0.033225	-0.529589
conv1	26	0.386193	30	2	0.011141	3.921553	0.102715	true	0.06418	0.055712	0.143184	5.779924	true	0.083353	0.10301	true	true	null	0.006109	0.536151
conv1	27	0.359935	75	0.25	0.002138	1.642503	0.101537	true	0.248787	0.090008	0.33224	1.234876	true	0.105717	0.135443	true	true	null	0.002818	-0.108557
conv1	28	0.570305	30	0	-0.002181	0.238708	0.123046	true	0.206631	0.219194	0.527703	1.65914	true	0.333239	0.406764	true	true	null	0.047409	-1.084715
conv1	29	0	0	2	0.034721	1.240257	0.096202	true	0.223019	0.150628	0.440735	-1.480597	false	0.121472	0.166061	false	true	null	0.001844	15.854401
conv1	30	0.293363	0	1.5	0.012395	3.596357	0.089638	true	0.126154	0.026907	0.164505	6.214319	true	0.031607	0.059456	true	true	null	0.000431	3.241999
conv1	31	0.022026	45	2	0.105623	0.070451	0.09659	true	0.09403	0.109605	0.278533	-0.656937	false	0.113407	0.194001	false	true	null	0.004171	16.396343
conv1	32	0.534811	15	0	-0.001369	0.409336	0.111006	true	0.140186	0.085356	0.229753	4.623257	true	0.176508	0.241607	true	true	null	0.018051	-0.776624
conv1	33	0.1925	90	0.25	0.000524	2.641212	0.089667	true	0.205236	0.088515	0.332267	-0.14388	false	0.020047	0.046339	true	true	null	-0.000467	0.855902
conv1	34	0	75	0	-0.000001	5.343097	0.103933	true	0.074626	0.071194	0.19523	-1.048209	false	0	0	false	true	null	0.035365	-0.616667
conv1	35	0.01956	105	2	0.14404	0.645152	0.103203	true	0.165611	0.07008	0.247231	-2.084052	false	0.021465	0.039448	false	true	null	0.015947	3.573153
conv1	36	0.039346	120	0	-0.000494	0.053677	0.092705	true	0.147175	0.095606	0.272654	-1.127838	false	0.086744	0.179068	false	true	null	0.015844	-0.580984
conv1	37	0	0	0	0	0	0.082747	true	0.068568	0.05528	0.160793	-1.240371	false	0.892199	0.980303	false	true	null	0.001973	-0.577913
conv1	38	0.171637	165	0	-0.008108	0.078081	0.12264	true	0.087309	0.081281	0.198846	1.037489	false	0.262451	0.384197	false	true	null	0.015588	-0.975776
conv1	39	0.207784	60	0	-0.00089	2.364129	0.128993	true	0.098099	0.096103	0.247537	1.141323	false	0.012512	0.028297	true	true	null	0.060513	-1.272567
conv1	40	0.473506	15	0.25	0.012151	1.251866	0.105511	true	0.122567	0.101215	0.243668	3.467254	true	0.389863	0.720865	false	true	null	0.027274	-0.239114
conv1	41	0	0	0	-0.00001	5.004131	0.087627	true	0.242333	0.09067	0.345935	-2.67268	false	0	0	false	true	null	0.015615	-0.520804
conv1	42	0.279962	45	0	-0.000644	3.283388	0.094055	true	0.133898	0.104041	0.253235	1.403912	true	0.024759	0.045049	true	true	null	0.030788	-0.564095
conv1	43	0.037798	15	2	0.121179	0.318286	0.117378	true	0.048913	0.035662	0.110747	-0.311683	false	0.03886	0.0605	false	true	null	0.021924	3.002584
conv1	44	0.578301	15	0	-0.001099	0.502597	0.099664	true	0.130254	0.067094	0.200291	6.677861	true	0.276139	0.617471	false	true	null	0.00693	-0.744576
conv1	45	0	0	0	-0.00002	0.983629	0.100254	true	0.091428	0.064769	0.192566	-1.411606	false	0.059858	0.086487	false	true	null	0.02438	-0.827632
conv1	46	0.227502	90	2	0.039651	2.150485	0.13535	true	0.141703	0.121206	0.31549	0.707875	false	0.028003	0.074545	true	true	null	0.022315	0.39979
conv1	47	0.003403	60	2	0.008081	0.588281	0.122874	true	0.066987	0.066169	0.157798	-0.960935	false	0.230321	0.38975	false	true	null	0.041676	-0.522301
conv1	48	1	0	0	0	0.000507	0.121261	true	0.122473	0.074685	0.178274	11.749667	true	0.260755	0.385669	true	true	null	0.024514	-0.765212
conv1	49	0	0	0	0	0	0.108068	true	0.072762	0.072009	0.178125	-1.010452	false	0.515924	0.830532	false	true	null	0.019189	-0.596918
conv1	50	0.347617	15	0.25	0.001496	1.703655	0.104054	true	0.100262	0.102938	0.273042	2.402953	true	0.144808	0.285821	true	true	null	0.018359	-0.451374
conv1	51	0.497424	165	0.25	0.000225	0.30689	0.093311	true	0.043792	0.021884	0.074101	20.729107	true	0.123178	0.298911	true	true	null	0.009918	-0.465824
conv1	52	0.33245	165	2	0.006822	3.640274	0.102103	true	0.097675	0.06837	0.150432	3.433863	true	0.052071	0.07611	true	true	null	0.016868	-0.322795
conv1	53	0.295678	90	2	0.016711	3.294454	0.095501	true	0.20598	0.152125	0.376148	0.589633	false	0.047257	0.088786	true	true	null	0.006523	1.007077
conv1	54	0.398492	0	1.5	0.000984	3.896289	0.111418	true	0.166526	0.096317	0.295357	2.408364	true	0.075084	0.12427	true	true	null	0.005067	-1.112757
conv1	55	0.284628	60	0	-0.000167	3.318181	0.104851	true	0.106537	0.080325	0.232278	2.217128	true	0.063174	0.100385	true	true	null	0.066187	-1.054883
conv1	56	0.186924	135	0	-0.000348	2.877241	0.105842	true	0.128918	0.090858	0.274721	0.638431	false	0.007512	0.010813	true	true	null	0.020188	-0.517269
conv1	57	0.626853	75	0	-0.000574	0.150869	0.106853	true	0.099061	0.068999	0.17099	7.649273	true	0.152479	0.217985	true	true	null	0.01807	-0.728393
conv1	58	0.357735	135	0	-0.000065	5.610014	0.098217	true	0.119755	0.143846	0.3487	1.654413	true	0.083135	0.123142	true	true	null	0.035369	-0.702795
conv1	59	0.248177	90	0	-0.000381	3.222099	0.109181	true	0.153792	0.100543	0.264586	0.938756	false	0.038298	0.061685	true	true	null	0.018601	-0.45443
conv1	60	0.217376	90	2	0.017325	3.174984	0.103531	true	0.127822	0.115218	0.31387	0.777259	false	0.023743	0.049108	true	true	null	0.011184	0.212159
conv1	61	0	0	0	-0.000003	1.053506	0.106748	true	0.105034	0.110713	0.292291	-0.948706	false	0.026491	0.043237	false	true	null	0.002994	-0.60751
conv1	62	0	15	0	-0.000238	1.531743	0.097917	true	0.145295	0.065816	0.217041	-2.207595	false	0.108549	0.239414	false	true	null	0.00102	-0.119548
conv1	63	0.312332	15	0	-0.001105	2.86195	0.103688	true	0.148741	0.092269	0.296204	1.772976	true	0.059034	0.079477	true	true	null	0.014839	-0.622161
conv2	0	0.831355	45	0	-0.002309	4.861965	0.025945	true	0.106964	0.093737	0.269411	7.727897	true	0.075636	0.10095	true	true	null	0.00776	-0.485323
conv2	1	0.132904	60	0	-0.001297	0.485515	0.015782	true	0.391942	0.392272	0.996572	-0.660352	false	0.081165	0.135416	false	true	null	0.00359	-0.640064
conv2	2	0.143069	165	2	0.017996	6.07367	0.020546	true	0.170082	0.214221	0.552095	-0.126101	false	0.086353	0.144964	false	true	null	-0.005514	1.368087
conv2	3	0.368151	120	0	-0.020633	2.562529	0.010921	true	0.290525	0.2512	0.718176	0.30902	false	0.294327	0.82568	false	true	null	-0.022308	0.014595
conv2	4	0.88905	105	0	-0.000593	2.103451	0.012256	true	0.164728	0.113032	0.323948	6.408096	true	0.074601	0.219408	true	true	null	0.0296	-0.907192
conv2	5	0.110658	30	0	-0.001747	1.214525	0.004657	true	0.58552	0.279315	0.919796	-1.700095	false	0.08199	0.151706	false	true	null	0.231458	-0.539283
conv2	6	0.007601	0	0	-0.000644	0.714241	0.015983	true	0.146922	0.104795	0.281011	-1.329464	false	0.063841	0.1324	false	true	null	-0.001441	0.546465
conv2	7	0.219675	135	0.25	0.000835	2.405078	0.003803	true	0.573634	0.356933	0.933388	-0.991667	false	0.290054	0.733464	false	true	null	0.183606	-0.530639
conv2	8	0.147288	30	0	-0.012896	1.020511	0.006342	true	0.411097	0.342247	0.817621	-0.770817	false	0.196506	0.316076	false	true	null	-0.084644	0.738894
conv2	9	0.304534	75	0	-0.000745	5.420515	0.007217	true	0.295414	0.258855	0.633806	0.035233	false	0.077423	0.234514	true	true	null	0.222672	-0.510829
conv2	10	0.831139	135	0.25	0.003658	1.313759	0.005693	true	0.522047	0.30851	0.950615	1.001888	false	0.738566	0.943873	false	true	null	0.139091	-0.271168
conv2	11	0.246065	105	0	-0.001994	5.656504	0.007559	true	0.283821	0.093979	0.403592	-0.401752	false	0.549302	1	false	true	null	0.217914	-0.501802
conv2	12	0.273733	165	0.25	0.001659	1.250455	0.008792	true	0.234699	0.345086	0.849472	0.113116	false	0.062908	0.18352	true	true	null	0.206672	-0.461949
conv2	13	0.572495	150	0	-0.001363	1.193101	0.013816	true	0.209536	0.146437	0.390259	2.478593	true	0.115726	0.234743	true	true	null	0.00091	-0.407454
conv2	14	0.883017	75	0	-0.007463	2.284706	0.012009	true	0.185251	0.245517	0.612778	2.842027	true	0.364523	0.90976	false	true	null	0.002297	-1.342813
conv2	15	0.289216	60	0.25	0.003014	1.353271	0.013678	true	0.340468	0.37934	0.972834	-0.135108	false	0.244876	0.77587	false	true	null	0.043626	-0.625063
conv2	16	0.275818	90	0	-0.009667	0.384335	0.01209	true	0.319908	0.381615	1	-0.115535	false	0.133327	0.182843	true	true	null	0.216935	-0.515906
conv2	17	0.168654	150	0	-0.00728	1.642164	0.008499	true	0.334887	0.29876	0.837737	-0.556407	false	0.110196	0.168756	false	true	null	-0.023159	0.349321
conv2	18	0.254822	150	0.25	0.000859	2.415795	0.012714	true	0.320932	0.360336	0.921844	-0.183468	false	0.067981	0.126374	true	true	null	-0.004804	0.178667
conv2	19	0.595382	75	0	-0.004606	4.782381	0.003615	true	0.379931	0.3245	0.671307	0.663949	false	0.171366	0.416063	true	true	null	-0.010537	0.238523
conv2	20	0.505795	15	0	-0.007232	3.592375	0.013983	true	0.361484	0.346459	0.830249	0.416532	false	0.164006	0.569447	false	true	null	-0.011389	0.114731
conv2	21	0.053499	105	2	0.001314	0.87073	0.020983	true	0.356566	0.33054	0.839122	-0.916884	false	0.47283	0.88021	false	true	null	0.017404	-0.2922
conv2	22	0.609645	15	0	-0.002064	2.955877	0.009206	true	0.116506	0.077025	0.203965	6.402318	true	0.048173	0.111037	true	true	null	-0.069236	0.670636
conv2	23	0.129149	15	0	-0.000557	0.76611	0.007714	true	0.410751	0.328969	0.879455	-0.856017	false	0.060353	0.195466	false	true	null	0.009177	-0.108802
conv2	24	0.242829	120	0	-0.005394	0.644539	0.011094	true	0.133497	0.079244	0.23525	1.379684	true	0.227654	0.715239	false	true	null	-0.014214	0.406755
conv2	25	0.75199	105	0	-0.005445	0.800593	0.011242	true	0.197657	0.12351	0.37755	4.488163	true	0.059126	0.139452	true	true	null	0.003048	-0.187349
conv2	26	0.241439	150	0	-0.002787	0.171429	0.008147	true	0.445722	0.328063	0.876742	-0.622694	false	0.458186	0.977289	false	true	null	0.181933	-0.444624
conv2	27	0.11225	0	0	-0.000255	0.582804	0.00747	true	0.585081	0.448175	0.999117	-1.055014	false	0.863423	1	false	true	null	0.023282	-0.388338
conv2	28	0.077256	0	0	-0.01559	0.294545	0.011208	true	0.369491	0.358494	0.85918	-0.815175	false	0.042945	0.081936	false	true	null	0.017107	-0.707831
conv2	29	0.273082	150	0	-0.000483	0.549053	0.013665	true	0.376175	0.252198	0.742706	-0.408779	false	0.598622	0.999832	false	true	null	-0.022526	0.362803
conv2	30	0.223506	15	0	-0.014692	2.092565	0.005988	true	0.333032	0.268376	0.691619	-0.408108	false	0.525244	1	false	true	null	0.006896	-1.916457
conv2	31	0.447631	30	0	-0.008831	2.073156	0.002409	true	0.35068	0.338694	0.834183	0.286248	false	0.23293	0.566353	false	true	null	0.198678	-0.523181
conv2	32	0.373062	165	0	-0.012703	2.897519	0.011205	true	0.304429	0.169393	0.577823	0.405172	false	0.239778	0.291219	true	true	null	0.006592	-0.970721
conv2	33	0.27523	30	0.25	0.001475	1.514214	0.006043	true	0.406025	0.389385	0.963267	-0.335901	false	0.300525	1	false	true	null	0.118492	-0.539111
conv2	34	0.202399	30	0.5	0.001026	1.203304	0.01336	true	0.108645	0.13166	0.341849	0.712085	false	0.302496	0.502161	false	true	null	-0.015041	0.39085
conv2	35	0.047447	15	0	-0.000151	1.704452	0.008305	true	0.153888	0.113424	0.304616	-0.938437	false	0.061045	0.120058	false	true	null	0.004547	-0.26915

End of preview. Expand in Data Studio

InceptionV1 tuning atlas — every unit, with its nulls

Quantitative per-unit orientation and curvature tuning for all 5,808 units of InceptionV1 (torchvision googlenet), each shipped with two null distributions.

Why this exists. The Distill Circuits thread established oriented-edge and curve detectors with feature visualizations and rendered tuning-curve widgets. None of it was published as numbers, no randomization control appears anywhere in the thread, and OpenAI Microscope — the visualization layer — has returned HTTP 503 since roughly January 2025. The founding rung of mechanistic interpretability is currently its least reproducible.

Fields (one JSON object per unit)

layer, unit, osi (orientation selectivity, 1 − circular variance, measured at that unit's preferred spatial frequency), pref_orientation_deg, pref_curvature, curve_selectivity_index, response_magnitude, and per-null: null_osi_{mean,std,max}, osi_z_vs_null, osi_exceeds_all_nulls, null_alive, shuffle_osi_{mean,max}, osi_exceeds_all_shuffles, shuffle_alive.

Two nulls, because the standard one is too weak

Random-init (Adebayo et al. 2018): same architecture, random weights.
Weight-shuffle: permute each trained kernel's weights within-channel — preserves the weight distribution and keeps activations alive.

Both are distributions over 5 seeds. The distinction matters:

	beats random-init	beats weight-shuffle
`conv1`	34/64	40/64
`inception5b`	100%	46%

A randomly-initialized network is nearly dead below inception3a (0% of units respond), so "beats the random-init null" is trivially true there. null_alive flags exactly where the weaker null is uninformative.

Honest headline

Trained top OSI 1.000 vs random-init 0.528 and shuffle 0.980 — the sharpest detectors are real. But the median conv1 unit sits at ~0.25: "conv1 is all Gabor filters" is too strong a reading, and the numbers say so where the pictures could not.

Method traps this cost us

Measure orientation tuning at each unit's preferred spatial frequency — averaging across frequencies inverted the conv1 result entirely.
Use a bounded curve index, not a ratio; a near-zero straight-line response produced "selectivities" of 2,000,000.
A dead null is not a null.

Caveat on unit indices

These are torchvision googlenet channel indices. They are not known to correspond to the unit numbering in the Distill articles, which refers to the lucid/TF-slim InceptionV1 checkpoint. Do not join these to those labels without first establishing the mapping.

Reproduce: scripts/inceptionv1_atlas.py (CPU, minutes). Interactive viewer and curriculum: spinning-up-in-mech-interp. Morgan Hough, Orthogonal Research and Education Lab (OREL).

Downloads last month: 21