Journal: Environmental Research Communications

Article Title: How to estimate carbon footprint when training deep learning models? A guide and review

doi: 10.1088/2515-7620/acf81b

Figure Lengend Snippet: Summary of the characteristics of the energy and CO2eq measurement tools. Wattmeters are not included in the table.

Article Snippet: Cumulator , The data of emission intensity is from Electricity Maps . Information is collected in the country_dataset_adjusted.csv file. The default value is 447 gCO2eq/kWh (average carbon intensity value in gCO2eq/kWh in the EU in 2018 Moro and Lonza ). .

Techniques: Modification

Journal: Environmental Research Communications

Article Title: How to estimate carbon footprint when training deep learning models? A guide and review

doi: 10.1088/2515-7620/acf81b

Figure Lengend Snippet: Estimation of energy consumption for CPUs.

Article Snippet: Cumulator , The data of emission intensity is from Electricity Maps . Information is collected in the country_dataset_adjusted.csv file. The default value is 447 gCO2eq/kWh (average carbon intensity value in gCO2eq/kWh in the EU in 2018 Moro and Lonza ). .

Techniques: Software

Journal: Environmental Research Communications

Article Title: How to estimate carbon footprint when training deep learning models? A guide and review

doi: 10.1088/2515-7620/acf81b

Figure Lengend Snippet: Estimation of energy consumption for GPUs.

Article Snippet: Cumulator , The data of emission intensity is from Electricity Maps . Information is collected in the country_dataset_adjusted.csv file. The default value is 447 gCO2eq/kWh (average carbon intensity value in gCO2eq/kWh in the EU in 2018 Moro and Lonza ). .

Techniques:

Journal: Environmental Research Communications

Article Title: How to estimate carbon footprint when training deep learning models? A guide and review

doi: 10.1088/2515-7620/acf81b

Figure Lengend Snippet: Estimation of energy consumption for memory.

Article Snippet: Cumulator , The data of emission intensity is from Electricity Maps . Information is collected in the country_dataset_adjusted.csv file. The default value is 447 gCO2eq/kWh (average carbon intensity value in gCO2eq/kWh in the EU in 2018 Moro and Lonza ). .

Techniques:

Journal: Environmental Research Communications

Article Title: How to estimate carbon footprint when training deep learning models? A guide and review

doi: 10.1088/2515-7620/acf81b

Figure Lengend Snippet: PUE values used in the different tools.

Article Snippet: Cumulator , The data of emission intensity is from Electricity Maps . Information is collected in the country_dataset_adjusted.csv file. The default value is 447 gCO2eq/kWh (average carbon intensity value in gCO2eq/kWh in the EU in 2018 Moro and Lonza ). .

Techniques:

Journal: Environmental Research Communications

Article Title: How to estimate carbon footprint when training deep learning models? A guide and review

doi: 10.1088/2515-7620/acf81b

Figure Lengend Snippet: Emission intensity used in the different tools.

Article Snippet: Cumulator , The data of emission intensity is from Electricity Maps . Information is collected in the country_dataset_adjusted.csv file. The default value is 447 gCO2eq/kWh (average carbon intensity value in gCO2eq/kWh in the EU in 2018 Moro and Lonza ). .

Techniques:

Journal: Environmental Research Communications

Article Title: How to estimate carbon footprint when training deep learning models? A guide and review

doi: 10.1088/2515-7620/acf81b

Figure Lengend Snippet: Requirements to run the tools.

Article Snippet: Cumulator , The data of emission intensity is from Electricity Maps . Information is collected in the country_dataset_adjusted.csv file. The default value is 447 gCO2eq/kWh (average carbon intensity value in gCO2eq/kWh in the EU in 2018 Moro and Lonza ). .

Techniques:

Journal: Environmental Research Communications

Article Title: How to estimate carbon footprint when training deep learning models? A guide and review

doi: 10.1088/2515-7620/acf81b

Figure Lengend Snippet: Results for the training of a digit classifier (experiment 1). All consumption values are in Wh. Carbon emissions are in gCO2e. For CodeCarbon and Eco2AI, (P) refers to the process tracking mode and (M) to the machine tracking mode.

Article Snippet: Cumulator , The data of emission intensity is from Electricity Maps . Information is collected in the country_dataset_adjusted.csv file. The default value is 447 gCO2eq/kWh (average carbon intensity value in gCO2eq/kWh in the EU in 2018 Moro and Lonza ). .

Techniques:

Journal: Environmental Research Communications

Article Title: How to estimate carbon footprint when training deep learning models? A guide and review

doi: 10.1088/2515-7620/acf81b

Figure Lengend Snippet: Results for the training of an image denoiser (experiment 2). All consumption values are in kWh. Carbon emissions are in gCO2e. The consumption indicated for Colab is extrapolated. An epoch was executed, the consumptions were obtained, and the values were extrapolated.

Article Snippet: Cumulator , The data of emission intensity is from Electricity Maps . Information is collected in the country_dataset_adjusted.csv file. The default value is 447 gCO2eq/kWh (average carbon intensity value in gCO2eq/kWh in the EU in 2018 Moro and Lonza ). .

Techniques:

Journal: Environmental Research Communications

Article Title: How to estimate carbon footprint when training deep learning models? A guide and review

doi: 10.1088/2515-7620/acf81b

Figure Lengend Snippet: Results running experiment 2 twice in parallel on Gemini-1: one process using trackers, the other without.

Article Snippet: Cumulator , The data of emission intensity is from Electricity Maps . Information is collected in the country_dataset_adjusted.csv file. The default value is 447 gCO2eq/kWh (average carbon intensity value in gCO2eq/kWh in the EU in 2018 Moro and Lonza ). .

Techniques:

Journal: The Journal of Neuroscience

Article Title: Evoked Response Strength in Primary Auditory Cortex Predicts Performance in a Spectro-Spatial Discrimination Task in Rats

doi: 10.1523/JNEUROSCI.0041-18.2019

Figure Lengend Snippet: Individual animal task acquisition through consecutive shaping stages. A, Learning curves for sound localization without distractor. Rat1 and rat2 were trained on pure tones (8 kHz tone sequence; blue marker), whereas rats 9–12 were trained on white noise (same duration and amplitude parameters; black marker) and transitioned to pure tones once stable performance was reached. A cumulative Weibull function, fitted to the mean performance of consecutive behavioral sessions, shows each animal's dynamic learning phase (shaded green). B, Stable localization performance for white noise (wn), 4, 8, and 16 kHz (the pure tones used in the final spectra-spatial discrimination task). Mean performance and 95% binomial CIs are shown. C, Individual performance for rats 9–12 during the transition from the localization paradigm (no distractor) to the final target-distractor discrimination paradigm. Mean session performance is shown for 16 kHz localization without distractor (circles), with distractor at a 10–20 dB lower amplitude (squares), and finally with distractor amplitude-matched to the target (crosses; 60 dB SPL). Performance falls only when the full-volume distractor is introduced and quickly recovers. Behavioral sessions with the 8 kHz target were interspersed in the training (data not shown for clarity). Rat1 and rat2 were transitioned to the discrimination paradigm using a different pure tone frequency (24 kHz), thus data are not shown.

Article Snippet: A cumulative Weibull function was fitted to performance over consecutive sessions of the localization paradigm (no distractor) to produce learning curves ( A ), with the dynamic learning phase defined as the range between the first and ninth decile of the function, as previously described by Mayrhofer et al. (2013) .

Techniques: Sequencing, Marker